Merge pull request #173 from THUDM/CogVideoX_dev

CogVideoX-5B config

.gitignore

@@ -1,4 +1,3 @@
-output/
 *__pycache__/
 samples*/
 runs/
@@ -7,3 +6,5 @@ master_ip
 logs/
 *.DS_Store
 .idea
+output*
+test*

README.md

@@ -22,7 +22,8 @@
 
 ## Update and News
 
-- 🔥🔥 **News**: ```2024/8/20```: [VEnhancer](https://github.com/Vchitect/VEnhancer) now supports enhancing videos generated by
+- 🔥🔥 **News**: ```2024/8/20```: [VEnhancer](https://github.com/Vchitect/VEnhancer) now supports enhancing videos
+  generated by
   CogVideoX, achieving higher resolution and higher quality video rendering. We welcome you to try it out by following
   the [tutorial](tools/venhancer/README_zh.md).
 - 🔥 **News**: ```2024/8/15```: The `SwissArmyTransformer` dependency in CogVideoX has been upgraded to `0.4.12`.
@@ -145,6 +146,9 @@ works have already been adapted for CogVideoX, and we invite everyone to use the
 + [Xorbits Inference](https://github.com/xorbitsai/inference): A powerful and comprehensive distributed inference
   framework, allowing you to easily deploy your own models or the latest cutting-edge open-source models with just one
   click.
++ [VideoSys](https://github.com/NUS-HPC-AI-Lab/VideoSys): VideoSys provides a user-friendly, high-performance
+  infrastructure for video generation, with full pipeline support and continuous integration of the latest models and
+  techniques.
 
 ## Project Structure
 

README_ja.md

@@ -130,6 +130,8 @@ CogVideoXは、[清影](https://chatglm.cn/video?fr=osm_cogvideox) と同源の
 
 + [Xorbits Inference](https://github.com/xorbitsai/inference):
   強力で包括的な分散推論フレームワークであり、ワンクリックで独自のモデルや最新のオープンソースモデルを簡単にデプロイできます。
++ [VideoSys](https://github.com/NUS-HPC-AI-Lab/VideoSys): VideoSysは、使いやすく高性能なビデオ生成インフラを提供し、最新のモデルや技術を継続的に統合しています。
+
 
 ## プロジェクト構造
 

README_zh.md

@@ -125,6 +125,7 @@ CogVideoX是 [清影](https://chatglm.cn/video?fr=osm_cogvideox) 同源的开源
 我们非常欢迎来自社区的贡献，并积极的贡献开源社区。以下作品已经对CogVideoX进行了适配，欢迎大家使用:
 
 + [Xorbits Inference](https://github.com/xorbitsai/inference): 性能强大且功能全面的分布式推理框架，轻松一键部署你自己的模型或内置的前沿开源模型。
++ [VideoSys](https://github.com/NUS-HPC-AI-Lab/VideoSys): VideoSys 提供了易用且高性能的视频生成基础设施，支持完整的管道，并持续集成最新的模型和技术。
 
 ## 完整项目代码结构
 

inference/cli_demo.py

@@ -10,31 +10,10 @@ Run the script:
 
 """
 
-import gc
 import argparse
-import tempfile
-from typing import Union, List
-
-import PIL
-import imageio
-import numpy as np
 import torch
 from diffusers import CogVideoXPipeline, CogVideoXDDIMScheduler
-
+from diffusers.utils import export_to_video
-def export_to_video_imageio(
-        video_frames: Union[List[np.ndarray], List[PIL.Image.Image]], output_video_path: str = None, fps: int = 8
-) -> str:
-    """
-    Export the video frames to a video file using imageio lib to Avoid "green screen" issue (for example CogVideoX)
-    """
-    if output_video_path is None:
-        output_video_path = tempfile.NamedTemporaryFile(suffix=".mp4").name
-    if isinstance(video_frames[0], PIL.Image.Image):
-        video_frames = [np.array(frame) for frame in video_frames]
-    with imageio.get_writer(output_video_path, fps=fps) as writer:
-        for frame in video_frames:
-            writer.append_data(frame)
-    return output_video_path
 
 
 def generate_video(
@@ -44,7 +23,7 @@ def generate_video(
         num_inference_steps: int = 50,
         guidance_scale: float = 6.0,
         num_videos_per_prompt: int = 1,
-        dtype: torch.dtype = torch.float16,
+        dtype: torch.dtype = torch.bfloat16,
 ):
     """
     Generates a video based on the given prompt and saves it to the specified path.
@@ -56,11 +35,11 @@ def generate_video(
     - num_inference_steps (int): Number of steps for the inference process. More steps can result in better quality.
     - guidance_scale (float): The scale for classifier-free guidance. Higher values can lead to better alignment with the prompt.
     - num_videos_per_prompt (int): Number of videos to generate per prompt.
-    - dtype (torch.dtype): The data type for computation (default is torch.float16).
+    - dtype (torch.dtype): The data type for computation (default is torch.bfloat16).
 
     """
 
-    # 1.  Load the pre-trained CogVideoX pipeline with the specified precision (float16).
+    # 1.  Load the pre-trained CogVideoX pipeline with the specified precision (bfloat16).
     # add device_map="balanced" in the from_pretrained function and remove the enable_model_cpu_offload()
     # function to use Multi GPUs.
 
@@ -71,39 +50,32 @@ def generate_video(
     # We recommend using `CogVideoXDDIMScheduler` for better results.
     pipe.scheduler = CogVideoXDDIMScheduler.from_config(pipe.scheduler.config, timestep_spacing="trailing")
 
-    # 3. Enable CPU offload for the model and reset the memory, enable tiling.
+    # 3. Enable CPU offload for the model, enable tiling.
     pipe.enable_model_cpu_offload()
-
+    pipe.vae.enable_tiling()
-    gc.collect()
-    torch.cuda.empty_cache()
-    torch.cuda.reset_accumulated_memory_stats()
-    torch.cuda.reset_peak_memory_stats()
-
-    # Using with diffusers branch `main` to enable tiling. This will cost ONLY 12GB GPU memory.
-    # pipe.vae.enable_tiling()
 
     # 4. Generate the video frames based on the prompt.
     # `num_frames` is the Number of frames to generate.
     # This is the default value for 6 seconds video and 8 fps,so 48 frames and will plus 1 frame for the first frame.
-    # for diffusers version `0.30.0`, this should be 48. and for `0.31.0` and after, this should be 49.
+    # for diffusers `0.30.1` and after version, this should be 49.
     video = pipe(
         prompt=prompt,
         num_videos_per_prompt=num_videos_per_prompt,  # Number of videos to generate per prompt
         num_inference_steps=num_inference_steps,  # Number of inference steps
-        num_frames=48, # Number of frames to generate，changed to 49 for diffusers version `0.31.0` and after.
+        num_frames=49,  # Number of frames to generate，changed to 49 for diffusers version `0.31.0` and after.
         guidance_scale=guidance_scale,  # Guidance scale for classifier-free guidance
         generator=torch.Generator().manual_seed(42),  # Set the seed for reproducibility
     ).frames[0]
 
-    # 5. Export the generated frames to a video file. fps must be 8
+    # 5. Export the generated frames to a video file. fps must be 8 for original video.
-    export_to_video_imageio(video, output_path, fps=8)
+    export_to_video(video, output_path, fps=8)
 
 
 if __name__ == "__main__":
     parser = argparse.ArgumentParser(description="Generate a video from a text prompt using CogVideoX")
     parser.add_argument("--prompt", type=str, required=True, help="The description of the video to be generated")
     parser.add_argument(
-        "--model_path", type=str, default="THUDM/CogVideoX-2b", help="The path of the pre-trained model to be used"
+        "--model_path", type=str, default="THUDM/CogVideoX-5b", help="The path of the pre-trained model to be used"
     )
     parser.add_argument(
         "--output_path", type=str, default="./output.mp4", help="The path where the generated video will be saved"
@@ -114,13 +86,15 @@ if __name__ == "__main__":
     parser.add_argument("--guidance_scale", type=float, default=6.0, help="The scale for classifier-free guidance")
     parser.add_argument("--num_videos_per_prompt", type=int, default=1, help="Number of videos to generate per prompt")
     parser.add_argument(
-        "--dtype", type=str, default="float16", help="The data type for computation (e.g., 'float16' or 'float32')"
+        "--dtype", type=str, default="bfloat16", help="The data type for computation (e.g., 'float16' or 'bfloat16')"
     )
 
     args = parser.parse_args()
 
-    # Convert dtype argument to torch.dtype, NOT suggest BF16.
+    # Convert dtype argument to torch.dtype.
-    dtype = torch.float16 if args.dtype == "float16" else torch.float32
+    # For CogVideoX-2B model, use torch.float16.
+    # For CogVideoX-5B model, use torch.bfloat16.
+    dtype = torch.float16 if args.dtype == "float16" else torch.bfloat16
 
     # main function to generate video.
     generate_video(

inference/cli_demo_quantization.py

@@ -0,0 +1,105 @@
+"""
+This script demonstrates how to generate a video from a text prompt using CogVideoX with 🤗Huggingface Diffusers Pipeline.
+
+Note:
+    This script requires the `diffusers>=0.30.1` and `torchao>=0.4.0` library to be installed.
+
+Run the script:
+    $ python cli_demo.py --prompt "A girl ridding a bike." --model_path THUDM/CogVideoX-2b
+
+In this script, we have only provided the script for testing and inference in INT8 for the entire process
+(including T5 Encoder, CogVideoX Transformer, VAE).
+You can use other functionalities provided by torchao to convert to other precisions.
+Please note that INT4 is not supported.
+"""
+import argparse
+
+import torch
+from diffusers import AutoencoderKLCogVideoX, CogVideoXTransformer3DModel, CogVideoXPipeline
+from diffusers.utils import export_to_video
+from transformers import T5EncoderModel
+
+# Make sure to install torchao>=0.4.0
+from torchao.quantization import quantize_, int8_weight_only
+
+
+def generate_video(
+        prompt: str,
+        model_path: str,
+        output_path: str = "./output.mp4",
+        num_inference_steps: int = 50,
+        guidance_scale: float = 6.0,
+        num_videos_per_prompt: int = 1,
+        dtype: torch.dtype = torch.bfloat16,
+):
+    """
+    Generates a video based on the given prompt and saves it to the specified path.
+
+    Parameters:
+    - prompt (str): The description of the video to be generated.
+    - model_path (str): The path of the pre-trained model to be used.
+    - output_path (str): The path where the generated video will be saved.
+    - num_inference_steps (int): Number of steps for the inference process. More steps can result in better quality.
+    - guidance_scale (float): The scale for classifier-free guidance. Higher values can lead to better alignment with the prompt.
+    - num_videos_per_prompt (int): Number of videos to generate per prompt.
+    - dtype (torch.dtype): The data type for computation (default is torch.bfloat16).
+
+    """
+
+    text_encoder = T5EncoderModel.from_pretrained(model_path, subfolder="text_encoder", torch_dtype=dtype)
+    quantize_(text_encoder, int8_weight_only())
+    transformer = CogVideoXTransformer3DModel.from_pretrained(model_path, subfolder="transformer",
+                                                              torch_dtype=dtype)
+    quantize_(transformer, int8_weight_only())
+    vae = AutoencoderKLCogVideoX.from_pretrained(model_path, subfolder="vae", torch_dtype=dtype)
+    quantize_(vae, int8_weight_only())
+    pipe = CogVideoXPipeline.from_pretrained(
+        model_path,
+        text_encoder=text_encoder,
+        transformer=transformer,
+        vae=vae,
+        torch_dtype=dtype,
+    )
+    pipe.enable_model_cpu_offload()
+    pipe.vae.enable_tiling()
+    video = pipe(
+        prompt=prompt,
+        num_videos_per_prompt=num_videos_per_prompt,
+        num_inference_steps=num_inference_steps,
+        num_frames=49,
+        guidance_scale=guidance_scale,
+        generator=torch.Generator().manual_seed(42),
+    ).frames[0]
+
+    export_to_video(video, output_path, fps=8)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="Generate a video from a text prompt using CogVideoX")
+    parser.add_argument("--prompt", type=str, required=True, help="The description of the video to be generated")
+    parser.add_argument(
+        "--model_path", type=str, default="THUDM/CogVideoX-5b", help="The path of the pre-trained model to be used"
+    )
+    parser.add_argument(
+        "--output_path", type=str, default="./output.mp4", help="The path where the generated video will be saved"
+    )
+    parser.add_argument(
+        "--num_inference_steps", type=int, default=50, help="Number of steps for the inference process"
+    )
+    parser.add_argument("--guidance_scale", type=float, default=6.0, help="The scale for classifier-free guidance")
+    parser.add_argument("--num_videos_per_prompt", type=int, default=1, help="Number of videos to generate per prompt")
+    parser.add_argument(
+        "--dtype", type=str, default="bfloat16", help="The data type for computation (e.g., 'float16' or 'bfloat16')"
+    )
+
+    args = parser.parse_args()
+    dtype = torch.float16 if args.dtype == "float16" else torch.bfloat16
+    generate_video(
+        prompt=args.prompt,
+        model_path=args.model_path,
+        output_path=args.output_path,
+        num_inference_steps=args.num_inference_steps,
+        guidance_scale=args.guidance_scale,
+        num_videos_per_prompt=args.num_videos_per_prompt,
+        dtype=dtype,
+    )

inference/gradio_composite_demo/README.md

@@ -0,0 +1,45 @@
+---
+title: CogVideoX-5B
+emoji: 🎥
+colorFrom: yellow
+colorTo: blue
+sdk: gradio
+sdk_version: 4.42.0
+suggested_hardware: a10g-large
+suggested_storage: large
+app_port: 7860
+app_file: app.py
+models:
+  - THUDM/CogVideoX-5b
+tags:
+  - cogvideox
+  - video-generation
+  - thudm
+short_description: Text-to-Video
+disable_embedding: false
+---
+
+# Gradio Composite Demo
+
+This Gradio demo integrates the CogVideoX-5B model, allowing you to perform video inference directly in your browser. It
+supports features like UpScale, RIFE, and other functionalities.
+
+## Environment Setup
+
+Set the following environment variables in your system:
+
++ OPENAI_API_KEY = your_api_key
++ OPENAI_BASE_URL= your_base_url
++ GRADIO_TEMP_DIR= gradio_tmp
+
+## Installation
+
+```bash
+pip install -r requirements.txt 
+```
+
+## Running the code
+
+```bash
+python gradio_web_demo.py
+```

inference/gradio_composite_demo/app.py

@@ -0,0 +1,317 @@
+import math
+import os
+import random
+import threading
+import time
+
+import gradio as gr
+import torch
+from diffusers import CogVideoXPipeline, CogVideoXDDIMScheduler,CogVideoXDPMScheduler
+from datetime import datetime, timedelta
+
+from diffusers.image_processor import VaeImageProcessor
+from openai import OpenAI
+import moviepy.editor as mp
+import utils
+from rife_model import load_rife_model, rife_inference_with_latents
+from huggingface_hub import hf_hub_download, snapshot_download
+
+device = "cuda" if torch.cuda.is_available() else "cpu"
+
+hf_hub_download(repo_id="ai-forever/Real-ESRGAN", filename="RealESRGAN_x4.pth", local_dir="model_real_esran")
+snapshot_download(repo_id="AlexWortega/RIFE", local_dir="model_rife")
+
+pipe = CogVideoXPipeline.from_pretrained("THUDM/CogVideoX-5b", torch_dtype=torch.bfloat16).to(device)
+pipe.scheduler = CogVideoXDPMScheduler.from_config(pipe.scheduler.config, timestep_spacing="trailing")
+
+pipe.transformer.to(memory_format=torch.channels_last)
+pipe.transformer = torch.compile(pipe.transformer, mode="max-autotune", fullgraph=True)
+
+os.makedirs("./output", exist_ok=True)
+os.makedirs("./gradio_tmp", exist_ok=True)
+
+upscale_model = utils.load_sd_upscale("model_real_esran/RealESRGAN_x4.pth", device)
+frame_interpolation_model = load_rife_model("model_rife")
+
+sys_prompt = """You are part of a team of bots that creates videos. You work with an assistant bot that will draw anything you say in square brackets.
+
+For example , outputting " a beautiful morning in the woods with the sun peaking through the trees " will trigger your partner bot to output an video of a forest morning , as described. You will be prompted by people looking to create detailed , amazing videos. The way to accomplish this is to take their short prompts and make them extremely detailed and descriptive.
+There are a few rules to follow:
+
+You will only ever output a single video description per user request.
+
+When modifications are requested , you should not simply make the description longer . You should refactor the entire description to integrate the suggestions.
+Other times the user will not want modifications , but instead want a new image . In this case , you should ignore your previous conversation with the user.
+
+Video descriptions must have the same num of words as examples below. Extra words will be ignored.
+"""
+
+
+def convert_prompt(prompt: str, retry_times: int = 3) -> str:
+    if not os.environ.get("OPENAI_API_KEY"):
+        return prompt
+    client = OpenAI()
+    text = prompt.strip()
+
+    for i in range(retry_times):
+        response = client.chat.completions.create(
+            messages=[
+                {"role": "system", "content": sys_prompt},
+                {
+                    "role": "user",
+                    "content": 'Create an imaginative video descriptive caption or modify an earlier caption for the user input : "a girl is on the beach"',
+                },
+                {
+                    "role": "assistant",
+                    "content": "A radiant woman stands on a deserted beach, arms outstretched, wearing a beige trench coat, white blouse, light blue jeans, and chic boots, against a backdrop of soft sky and sea. Moments later, she is seen mid-twirl, arms exuberant, with the lighting suggesting dawn or dusk. Then, she runs along the beach, her attire complemented by an off-white scarf and black ankle boots, the tranquil sea behind her. Finally, she holds a paper airplane, her pose reflecting joy and freedom, with the ocean's gentle waves and the sky's soft pastel hues enhancing the serene ambiance.",
+                },
+                {
+                    "role": "user",
+                    "content": 'Create an imaginative video descriptive caption or modify an earlier caption for the user input : "A man jogging on a football field"',
+                },
+                {
+                    "role": "assistant",
+                    "content": "A determined man in athletic attire, including a blue long-sleeve shirt, black shorts, and blue socks, jogs around a snow-covered soccer field, showcasing his solitary exercise in a quiet, overcast setting. His long dreadlocks, focused expression, and the serene winter backdrop highlight his dedication to fitness. As he moves, his attire, consisting of a blue sports sweatshirt, black athletic pants, gloves, and sneakers, grips the snowy ground. He is seen running past a chain-link fence enclosing the playground area, with a basketball hoop and children's slide, suggesting a moment of solitary exercise amidst the empty field.",
+                },
+                {
+                    "role": "user",
+                    "content": 'Create an imaginative video descriptive caption or modify an earlier caption for the user input : " A woman is dancing, HD footage, close-up"',
+                },
+                {
+                    "role": "assistant",
+                    "content": "A young woman with her hair in an updo and wearing a teal hoodie stands against a light backdrop, initially looking over her shoulder with a contemplative expression. She then confidently makes a subtle dance move, suggesting rhythm and movement. Next, she appears poised and focused, looking directly at the camera. Her expression shifts to one of introspection as she gazes downward slightly. Finally, she dances with confidence, her left hand over her heart, symbolizing a poignant moment, all while dressed in the same teal hoodie against a plain, light-colored background.",
+                },
+                {
+                    "role": "user",
+                    "content": f'Create an imaginative video descriptive caption or modify an earlier caption in ENGLISH for the user input: "{text}"',
+                },
+            ],
+            model="glm-4-0520",
+            temperature=0.01,
+            top_p=0.7,
+            stream=False,
+            max_tokens=200,
+        )
+        if response.choices:
+            return response.choices[0].message.content
+    return prompt
+
+
+def infer(
+        prompt: str,
+        num_inference_steps: int,
+        guidance_scale: float,
+        seed: int = -1,
+        #progress=gr.Progress(track_tqdm=True),
+):
+    if seed == -1:
+        seed = random.randint(0, 2 ** 8 - 1)
+    video_pt = pipe(
+        prompt=prompt,
+        num_videos_per_prompt=1,
+        num_inference_steps=num_inference_steps,
+        num_frames=49,
+        use_dynamic_cfg=True,
+        output_type="pt",
+        guidance_scale=guidance_scale,
+        generator=torch.Generator(device="cpu").manual_seed(seed),
+    ).frames
+
+    return (video_pt, seed)
+
+
+def convert_to_gif(video_path):
+    clip = mp.VideoFileClip(video_path)
+    clip = clip.set_fps(8)
+    clip = clip.resize(height=240)
+    gif_path = video_path.replace(".mp4", ".gif")
+    clip.write_gif(gif_path, fps=8)
+    return gif_path
+
+
+def delete_old_files():
+    while True:
+        now = datetime.now()
+        cutoff = now - timedelta(minutes=10)
+        directories = ["./output", "./gradio_tmp"]
+
+        for directory in directories:
+            for filename in os.listdir(directory):
+                file_path = os.path.join(directory, filename)
+                if os.path.isfile(file_path):
+                    file_mtime = datetime.fromtimestamp(os.path.getmtime(file_path))
+                    if file_mtime < cutoff:
+                        os.remove(file_path)
+        time.sleep(600)
+
+
+threading.Thread(target=delete_old_files, daemon=True).start()
+
+with gr.Blocks() as demo:
+    gr.Markdown("""
+           <div style="text-align: center; font-size: 32px; font-weight: bold; margin-bottom: 20px;">
+               CogVideoX-5B Huggingface Space🤗
+           </div>
+           <div style="text-align: center;">
+               <a href="https://huggingface.co/THUDM/CogVideoX-5B">🤗 5B Model Hub</a> |
+               <a href="https://github.com/THUDM/CogVideo">🌐 Github</a> |
+               <a href="https://arxiv.org/pdf/2408.06072">📜 arxiv </a>
+           </div>
+
+           <div style="text-align: center; font-size: 15px; font-weight: bold; color: red; margin-bottom: 20px;">
+            ⚠️ This demo is for academic research and experiential use only. 
+            </div>
+           """)
+    with gr.Row():
+        with gr.Column():
+            prompt = gr.Textbox(label="Prompt (Less than 200 Words)", placeholder="Enter your prompt here", lines=5)
+
+            with gr.Row():
+                gr.Markdown(
+                    "✨Upon pressing the enhanced prompt button, we will use [GLM-4 Model](https://github.com/THUDM/GLM-4) to polish the prompt and overwrite the original one."
+                )
+                enhance_button = gr.Button("✨ Enhance Prompt(Optional)")
+
+            gr.Markdown(
+                "<span style='color:red; font-weight:bold;'>For the CogVideoX-5B model, 50 steps will take approximately 120 seconds.</span>"
+            )
+
+            with gr.Group():
+                with gr.Column():
+                    with gr.Row():
+                        seed_param = gr.Number(
+                            label="Inference Seed (Enter a positive number, -1 for random)", value=-1
+                        )
+                    with gr.Row():
+                        enable_scale = gr.Checkbox(label="Super-Resolution (720 × 480 -> 1440 × 960)", value=False)
+                        enable_rife = gr.Checkbox(label="Frame Interpolation (8fps -> 16fps)", value=False)
+                    gr.Markdown(
+                        "✨In this demo, we use [RIFE](https://github.com/hzwer/ECCV2022-RIFE) for frame interpolation and [Real-ESRGAN](https://github.com/xinntao/Real-ESRGAN) for upscaling(Super-Resolution).<br>&nbsp;&nbsp;&nbsp;&nbsp;The entire process is based on open-source solutions."
+                    )
+
+            generate_button = gr.Button("🎬 Generate Video")
+
+        with gr.Column():
+            video_output = gr.Video(label="CogVideoX Generate Video", width=720, height=480)
+            with gr.Row():
+                download_video_button = gr.File(label="📥 Download Video", visible=False)
+                download_gif_button = gr.File(label="📥 Download GIF", visible=False)
+                seed_text = gr.Number(label="Seed Used for Video Generation", visible=False)
+
+    gr.Markdown("""
+    <table border="0" style="width: 100%; text-align: left; margin-top: 20px;">
+        <div style="text-align: center; font-size: 32px; font-weight: bold; margin-bottom: 20px;">
+            🎥 Video Gallery
+        </div>
+        <tr>
+            <td style="width: 25%; vertical-align: top; font-size: 0.9em;">
+                <p>A garden comes to life as a kaleidoscope of butterflies flutters amidst the blossoms, their delicate wings casting shadows on the petals below. In the background, a grand fountain cascades water with a gentle splendor, its rhythmic sound providing a soothing backdrop. Beneath the cool shade of a mature tree, a solitary wooden chair invites solitude and reflection, its smooth surface worn by the touch of countless visitors seeking a moment of tranquility in nature's embrace.</p>
+            </td>
+            <td style="width: 25%; vertical-align: top;">
+                <video src="https://github.com/user-attachments/assets/cf5953ea-96d3-48fd-9907-c4708752c714" width="100%" controls autoplay loop></video>
+            </td>
+            <td style="width: 25%; vertical-align: top; font-size: 0.9em;">
+                <p>A small boy, head bowed and determination etched on his face, sprints through the torrential downpour as lightning crackles and thunder rumbles in the distance. The relentless rain pounds the ground, creating a chaotic dance of water droplets that mirror the dramatic sky's anger. In the far background, the silhouette of a cozy home beckons, a faint beacon of safety and warmth amidst the fierce weather. The scene is one of perseverance and the unyielding spirit of a child braving the elements.</p>
+            </td>
+            <td style="width: 25%; vertical-align: top;">
+                <video src="https://github.com/user-attachments/assets/fe0a78e6-b669-4800-8cf0-b5f9b5145b52" width="100%" controls autoplay loop></video>
+            </td>
+        </tr>
+        <tr>
+            <td style="width: 25%; vertical-align: top; font-size: 0.9em;">
+                <p>A suited astronaut, with the red dust of Mars clinging to their boots, reaches out to shake hands with an alien being, their skin a shimmering blue, under the pink-tinged sky of the fourth planet. In the background, a sleek silver rocket, a beacon of human ingenuity, stands tall, its engines powered down, as the two representatives of different worlds exchange a historic greeting amidst the desolate beauty of the Martian landscape.</p>
+            </td>
+            <td style="width: 25%; vertical-align: top;">
+                <video src="https://github.com/user-attachments/assets/c182f606-8f8c-421d-b414-8487070fcfcb" width="100%" controls autoplay loop></video>
+            </td>
+            <td style="width: 25%; vertical-align: top; font-size: 0.9em;">
+                <p>An elderly gentleman, with a serene expression, sits at the water's edge, a steaming cup of tea by his side. He is engrossed in his artwork, brush in hand, as he renders an oil painting on a canvas that's propped up against a small, weathered table. The sea breeze whispers through his silver hair, gently billowing his loose-fitting white shirt, while the salty air adds an intangible element to his masterpiece in progress. The scene is one of tranquility and inspiration, with the artist's canvas capturing the vibrant hues of the setting sun reflecting off the tranquil sea.</p>
+            </td>
+            <td style="width: 25%; vertical-align: top;">
+                <video src="https://github.com/user-attachments/assets/7db2bbce-194d-434d-a605-350254b6c298" width="100%" controls autoplay loop></video>
+            </td>
+        </tr>
+        <tr>
+            <td style="width: 25%; vertical-align: top; font-size: 0.9em;">
+                <p>In a dimly lit bar, purplish light bathes the face of a mature man, his eyes blinking thoughtfully as he ponders in close-up, the background artfully blurred to focus on his introspective expression, the ambiance of the bar a mere suggestion of shadows and soft lighting.</p>
+            </td>
+            <td style="width: 25%; vertical-align: top;">
+                <video src="https://github.com/user-attachments/assets/62b01046-8cab-44cc-bd45-4d965bb615ec" width="100%" controls autoplay loop></video>
+            </td>
+            <td style="width: 25%; vertical-align: top; font-size: 0.9em;">
+                <p>A golden retriever, sporting sleek black sunglasses, with its lengthy fur flowing in the breeze, sprints playfully across a rooftop terrace, recently refreshed by a light rain. The scene unfolds from a distance, the dog's energetic bounds growing larger as it approaches the camera, its tail wagging with unrestrained joy, while droplets of water glisten on the concrete behind it. The overcast sky provides a dramatic backdrop, emphasizing the vibrant golden coat of the canine as it dashes towards the viewer.</p>
+            </td>
+            <td style="width: 25%; vertical-align: top;">
+                <video src="https://github.com/user-attachments/assets/d78e552a-4b3f-4b81-ac3f-3898079554f6" width="100%" controls autoplay loop></video>
+            </td>
+        </tr>
+        <tr>
+            <td style="width: 25%; vertical-align: top; font-size: 0.9em;">
+                <p>On a brilliant sunny day, the lakeshore is lined with an array of willow trees, their slender branches swaying gently in the soft breeze. The tranquil surface of the lake reflects the clear blue sky, while several elegant swans glide gracefully through the still water, leaving behind delicate ripples that disturb the mirror-like quality of the lake. The scene is one of serene beauty, with the willows' greenery providing a picturesque frame for the peaceful avian visitors.</p>
+            </td>
+            <td style="width: 25%; vertical-align: top;">
+                <video src="https://github.com/user-attachments/assets/30894f12-c741-44a2-9e6e-ddcacc231e5b" width="100%" controls autoplay loop></video>
+            </td>
+            <td style="width: 25%; vertical-align: top; font-size: 0.9em;">
+                <p>A Chinese mother, draped in a soft, pastel-colored robe, gently rocks back and forth in a cozy rocking chair positioned in the tranquil setting of a nursery. The dimly lit bedroom is adorned with whimsical mobiles dangling from the ceiling, casting shadows that dance on the walls. Her baby, swaddled in a delicate, patterned blanket, rests against her chest, the child's earlier cries now replaced by contented coos as the mother's soothing voice lulls the little one to sleep. The scent of lavender fills the air, adding to the serene atmosphere, while a warm, orange glow from a nearby nightlight illuminates the scene with a gentle hue, capturing a moment of tender love and comfort.</p>
+            </td>
+            <td style="width: 25%; vertical-align: top;">
+                <video src="https://github.com/user-attachments/assets/926575ca-7150-435b-a0ff-4900a963297b" width="100%" controls autoplay loop></video>
+            </td>
+        </tr>
+    </table>
+        """)
+
+
+    def generate(prompt,
+                 seed_value,
+                 scale_status,
+                 rife_status, 
+                 progress=gr.Progress(track_tqdm=True)
+                ):
+
+        latents, seed = infer(
+            prompt,
+            num_inference_steps=50,  # NOT Changed
+            guidance_scale=7.0,  # NOT Changed
+            seed=seed_value,
+            #progress=progress,
+        )
+        if scale_status:
+            latents = utils.upscale_batch_and_concatenate(upscale_model, latents, device)
+        if rife_status:
+            latents = rife_inference_with_latents(frame_interpolation_model, latents)
+
+        batch_size = latents.shape[0]
+        batch_video_frames = []
+        for batch_idx in range(batch_size):
+            pt_image = latents[batch_idx]
+            pt_image = torch.stack([pt_image[i] for i in range(pt_image.shape[0])])
+
+            image_np = VaeImageProcessor.pt_to_numpy(pt_image)
+            image_pil = VaeImageProcessor.numpy_to_pil(image_np)
+            batch_video_frames.append(image_pil)
+
+        video_path = utils.save_video(batch_video_frames[0], fps=math.ceil((len(batch_video_frames[0]) - 1) / 6))
+        video_update = gr.update(visible=True, value=video_path)
+        gif_path = convert_to_gif(video_path)
+        gif_update = gr.update(visible=True, value=gif_path)
+        seed_update = gr.update(visible=True, value=seed)
+
+        return video_path, video_update, gif_update, seed_update
+
+
+    def enhance_prompt_func(prompt):
+        return convert_prompt(prompt, retry_times=1)
+
+
+    generate_button.click(
+        generate,
+        inputs=[prompt, seed_param, enable_scale, enable_rife],
+        outputs=[video_output, download_video_button, download_gif_button, seed_text],
+    )
+
+    enhance_button.click(enhance_prompt_func, inputs=[prompt], outputs=[prompt])
+
+if __name__ == "__main__":
+    demo.launch()

inference/gradio_composite_demo/requirements.txt

@@ -0,0 +1,21 @@
+spaces==0.29.3
+safetensors>=0.4.4
+spandrel>=0.3.4
+tqdm>=4.66.5
+opencv-python>=4.10.0.84
+scikit-video>=1.1.11
+diffusers>=0.30.1
+transformers>=4.44.0
+accelerate>=0.33.0
+sentencepiece>=0.2.0
+SwissArmyTransformer>=0.4.12
+numpy==1.26.0
+torch>=2.4.0
+torchvision>=0.19.0
+gradio>=4.42.0
+streamlit>=1.37.1
+imageio==2.34.2
+imageio-ffmpeg==0.5.1
+openai>=1.42.0
+moviepy==1.0.3
+pillow==9.5.0

inference/gradio_composite_demo/rife/IFNet.py

@@ -0,0 +1,123 @@
+from .refine import *
+
+
+def deconv(in_planes, out_planes, kernel_size=4, stride=2, padding=1):
+    return nn.Sequential(
+        torch.nn.ConvTranspose2d(in_channels=in_planes, out_channels=out_planes, kernel_size=4, stride=2, padding=1),
+        nn.PReLU(out_planes),
+    )
+
+
+def conv(in_planes, out_planes, kernel_size=3, stride=1, padding=1, dilation=1):
+    return nn.Sequential(
+        nn.Conv2d(
+            in_planes,
+            out_planes,
+            kernel_size=kernel_size,
+            stride=stride,
+            padding=padding,
+            dilation=dilation,
+            bias=True,
+        ),
+        nn.PReLU(out_planes),
+    )
+
+
+class IFBlock(nn.Module):
+    def __init__(self, in_planes, c=64):
+        super(IFBlock, self).__init__()
+        self.conv0 = nn.Sequential(
+            conv(in_planes, c // 2, 3, 2, 1),
+            conv(c // 2, c, 3, 2, 1),
+        )
+        self.convblock = nn.Sequential(
+            conv(c, c),
+            conv(c, c),
+            conv(c, c),
+            conv(c, c),
+            conv(c, c),
+            conv(c, c),
+            conv(c, c),
+            conv(c, c),
+        )
+        self.lastconv = nn.ConvTranspose2d(c, 5, 4, 2, 1)
+
+    def forward(self, x, flow, scale):
+        if scale != 1:
+            x = F.interpolate(x, scale_factor=1.0 / scale, mode="bilinear", align_corners=False)
+        if flow != None:
+            flow = F.interpolate(flow, scale_factor=1.0 / scale, mode="bilinear", align_corners=False) * 1.0 / scale
+            x = torch.cat((x, flow), 1)
+        x = self.conv0(x)
+        x = self.convblock(x) + x
+        tmp = self.lastconv(x)
+        tmp = F.interpolate(tmp, scale_factor=scale * 2, mode="bilinear", align_corners=False)
+        flow = tmp[:, :4] * scale * 2
+        mask = tmp[:, 4:5]
+        return flow, mask
+
+
+class IFNet(nn.Module):
+    def __init__(self):
+        super(IFNet, self).__init__()
+        self.block0 = IFBlock(6, c=240)
+        self.block1 = IFBlock(13 + 4, c=150)
+        self.block2 = IFBlock(13 + 4, c=90)
+        self.block_tea = IFBlock(16 + 4, c=90)
+        self.contextnet = Contextnet()
+        self.unet = Unet()
+
+    def forward(self, x, scale=[4, 2, 1], timestep=0.5):
+        img0 = x[:, :3]
+        img1 = x[:, 3:6]
+        gt = x[:, 6:]  # In inference time, gt is None
+        flow_list = []
+        merged = []
+        mask_list = []
+        warped_img0 = img0
+        warped_img1 = img1
+        flow = None
+        loss_distill = 0
+        stu = [self.block0, self.block1, self.block2]
+        for i in range(3):
+            if flow != None:
+                flow_d, mask_d = stu[i](
+                    torch.cat((img0, img1, warped_img0, warped_img1, mask), 1), flow, scale=scale[i]
+                )
+                flow = flow + flow_d
+                mask = mask + mask_d
+            else:
+                flow, mask = stu[i](torch.cat((img0, img1), 1), None, scale=scale[i])
+            mask_list.append(torch.sigmoid(mask))
+            flow_list.append(flow)
+            warped_img0 = warp(img0, flow[:, :2])
+            warped_img1 = warp(img1, flow[:, 2:4])
+            merged_student = (warped_img0, warped_img1)
+            merged.append(merged_student)
+        if gt.shape[1] == 3:
+            flow_d, mask_d = self.block_tea(
+                torch.cat((img0, img1, warped_img0, warped_img1, mask, gt), 1), flow, scale=1
+            )
+            flow_teacher = flow + flow_d
+            warped_img0_teacher = warp(img0, flow_teacher[:, :2])
+            warped_img1_teacher = warp(img1, flow_teacher[:, 2:4])
+            mask_teacher = torch.sigmoid(mask + mask_d)
+            merged_teacher = warped_img0_teacher * mask_teacher + warped_img1_teacher * (1 - mask_teacher)
+        else:
+            flow_teacher = None
+            merged_teacher = None
+        for i in range(3):
+            merged[i] = merged[i][0] * mask_list[i] + merged[i][1] * (1 - mask_list[i])
+            if gt.shape[1] == 3:
+                loss_mask = (
+                    ((merged[i] - gt).abs().mean(1, True) > (merged_teacher - gt).abs().mean(1, True) + 0.01)
+                    .float()
+                    .detach()
+                )
+                loss_distill += (((flow_teacher.detach() - flow_list[i]) ** 2).mean(1, True) ** 0.5 * loss_mask).mean()
+        c0 = self.contextnet(img0, flow[:, :2])
+        c1 = self.contextnet(img1, flow[:, 2:4])
+        tmp = self.unet(img0, img1, warped_img0, warped_img1, mask, flow, c0, c1)
+        res = tmp[:, :3] * 2 - 1
+        merged[2] = torch.clamp(merged[2] + res, 0, 1)
+        return flow_list, mask_list[2], merged, flow_teacher, merged_teacher, loss_distill

inference/gradio_composite_demo/rife/IFNet_2R.py

@@ -0,0 +1,123 @@
+from .refine_2R import *
+
+
+def deconv(in_planes, out_planes, kernel_size=4, stride=2, padding=1):
+    return nn.Sequential(
+        torch.nn.ConvTranspose2d(in_channels=in_planes, out_channels=out_planes, kernel_size=4, stride=2, padding=1),
+        nn.PReLU(out_planes),
+    )
+
+
+def conv(in_planes, out_planes, kernel_size=3, stride=1, padding=1, dilation=1):
+    return nn.Sequential(
+        nn.Conv2d(
+            in_planes,
+            out_planes,
+            kernel_size=kernel_size,
+            stride=stride,
+            padding=padding,
+            dilation=dilation,
+            bias=True,
+        ),
+        nn.PReLU(out_planes),
+    )
+
+
+class IFBlock(nn.Module):
+    def __init__(self, in_planes, c=64):
+        super(IFBlock, self).__init__()
+        self.conv0 = nn.Sequential(
+            conv(in_planes, c // 2, 3, 1, 1),
+            conv(c // 2, c, 3, 2, 1),
+        )
+        self.convblock = nn.Sequential(
+            conv(c, c),
+            conv(c, c),
+            conv(c, c),
+            conv(c, c),
+            conv(c, c),
+            conv(c, c),
+            conv(c, c),
+            conv(c, c),
+        )
+        self.lastconv = nn.ConvTranspose2d(c, 5, 4, 2, 1)
+
+    def forward(self, x, flow, scale):
+        if scale != 1:
+            x = F.interpolate(x, scale_factor=1.0 / scale, mode="bilinear", align_corners=False)
+        if flow != None:
+            flow = F.interpolate(flow, scale_factor=1.0 / scale, mode="bilinear", align_corners=False) * 1.0 / scale
+            x = torch.cat((x, flow), 1)
+        x = self.conv0(x)
+        x = self.convblock(x) + x
+        tmp = self.lastconv(x)
+        tmp = F.interpolate(tmp, scale_factor=scale, mode="bilinear", align_corners=False)
+        flow = tmp[:, :4] * scale
+        mask = tmp[:, 4:5]
+        return flow, mask
+
+
+class IFNet(nn.Module):
+    def __init__(self):
+        super(IFNet, self).__init__()
+        self.block0 = IFBlock(6, c=240)
+        self.block1 = IFBlock(13 + 4, c=150)
+        self.block2 = IFBlock(13 + 4, c=90)
+        self.block_tea = IFBlock(16 + 4, c=90)
+        self.contextnet = Contextnet()
+        self.unet = Unet()
+
+    def forward(self, x, scale=[4, 2, 1], timestep=0.5):
+        img0 = x[:, :3]
+        img1 = x[:, 3:6]
+        gt = x[:, 6:]  # In inference time, gt is None
+        flow_list = []
+        merged = []
+        mask_list = []
+        warped_img0 = img0
+        warped_img1 = img1
+        flow = None
+        loss_distill = 0
+        stu = [self.block0, self.block1, self.block2]
+        for i in range(3):
+            if flow != None:
+                flow_d, mask_d = stu[i](
+                    torch.cat((img0, img1, warped_img0, warped_img1, mask), 1), flow, scale=scale[i]
+                )
+                flow = flow + flow_d
+                mask = mask + mask_d
+            else:
+                flow, mask = stu[i](torch.cat((img0, img1), 1), None, scale=scale[i])
+            mask_list.append(torch.sigmoid(mask))
+            flow_list.append(flow)
+            warped_img0 = warp(img0, flow[:, :2])
+            warped_img1 = warp(img1, flow[:, 2:4])
+            merged_student = (warped_img0, warped_img1)
+            merged.append(merged_student)
+        if gt.shape[1] == 3:
+            flow_d, mask_d = self.block_tea(
+                torch.cat((img0, img1, warped_img0, warped_img1, mask, gt), 1), flow, scale=1
+            )
+            flow_teacher = flow + flow_d
+            warped_img0_teacher = warp(img0, flow_teacher[:, :2])
+            warped_img1_teacher = warp(img1, flow_teacher[:, 2:4])
+            mask_teacher = torch.sigmoid(mask + mask_d)
+            merged_teacher = warped_img0_teacher * mask_teacher + warped_img1_teacher * (1 - mask_teacher)
+        else:
+            flow_teacher = None
+            merged_teacher = None
+        for i in range(3):
+            merged[i] = merged[i][0] * mask_list[i] + merged[i][1] * (1 - mask_list[i])
+            if gt.shape[1] == 3:
+                loss_mask = (
+                    ((merged[i] - gt).abs().mean(1, True) > (merged_teacher - gt).abs().mean(1, True) + 0.01)
+                    .float()
+                    .detach()
+                )
+                loss_distill += (((flow_teacher.detach() - flow_list[i]) ** 2).mean(1, True) ** 0.5 * loss_mask).mean()
+        c0 = self.contextnet(img0, flow[:, :2])
+        c1 = self.contextnet(img1, flow[:, 2:4])
+        tmp = self.unet(img0, img1, warped_img0, warped_img1, mask, flow, c0, c1)
+        res = tmp[:, :3] * 2 - 1
+        merged[2] = torch.clamp(merged[2] + res, 0, 1)
+        return flow_list, mask_list[2], merged, flow_teacher, merged_teacher, loss_distill

inference/gradio_composite_demo/rife/IFNet_HDv3.py

@@ -0,0 +1,138 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from .warplayer import warp
+
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+
+def conv(in_planes, out_planes, kernel_size=3, stride=1, padding=1, dilation=1):
+    return nn.Sequential(
+        nn.Conv2d(
+            in_planes,
+            out_planes,
+            kernel_size=kernel_size,
+            stride=stride,
+            padding=padding,
+            dilation=dilation,
+            bias=True,
+        ),
+        nn.PReLU(out_planes),
+    )
+
+
+def conv_bn(in_planes, out_planes, kernel_size=3, stride=1, padding=1, dilation=1):
+    return nn.Sequential(
+        nn.Conv2d(
+            in_planes,
+            out_planes,
+            kernel_size=kernel_size,
+            stride=stride,
+            padding=padding,
+            dilation=dilation,
+            bias=False,
+        ),
+        nn.BatchNorm2d(out_planes),
+        nn.PReLU(out_planes),
+    )
+
+
+class IFBlock(nn.Module):
+    def __init__(self, in_planes, c=64):
+        super(IFBlock, self).__init__()
+        self.conv0 = nn.Sequential(
+            conv(in_planes, c // 2, 3, 2, 1),
+            conv(c // 2, c, 3, 2, 1),
+        )
+        self.convblock0 = nn.Sequential(conv(c, c), conv(c, c))
+        self.convblock1 = nn.Sequential(conv(c, c), conv(c, c))
+        self.convblock2 = nn.Sequential(conv(c, c), conv(c, c))
+        self.convblock3 = nn.Sequential(conv(c, c), conv(c, c))
+        self.conv1 = nn.Sequential(
+            nn.ConvTranspose2d(c, c // 2, 4, 2, 1),
+            nn.PReLU(c // 2),
+            nn.ConvTranspose2d(c // 2, 4, 4, 2, 1),
+        )
+        self.conv2 = nn.Sequential(
+            nn.ConvTranspose2d(c, c // 2, 4, 2, 1),
+            nn.PReLU(c // 2),
+            nn.ConvTranspose2d(c // 2, 1, 4, 2, 1),
+        )
+
+    def forward(self, x, flow, scale=1):
+        x = F.interpolate(
+            x, scale_factor=1.0 / scale, mode="bilinear", align_corners=False, recompute_scale_factor=False
+        )
+        flow = (
+            F.interpolate(
+                flow, scale_factor=1.0 / scale, mode="bilinear", align_corners=False, recompute_scale_factor=False
+            )
+            * 1.0
+            / scale
+        )
+        feat = self.conv0(torch.cat((x, flow), 1))
+        feat = self.convblock0(feat) + feat
+        feat = self.convblock1(feat) + feat
+        feat = self.convblock2(feat) + feat
+        feat = self.convblock3(feat) + feat
+        flow = self.conv1(feat)
+        mask = self.conv2(feat)
+        flow = (
+            F.interpolate(flow, scale_factor=scale, mode="bilinear", align_corners=False, recompute_scale_factor=False)
+            * scale
+        )
+        mask = F.interpolate(
+            mask, scale_factor=scale, mode="bilinear", align_corners=False, recompute_scale_factor=False
+        )
+        return flow, mask
+
+
+class IFNet(nn.Module):
+    def __init__(self):
+        super(IFNet, self).__init__()
+        self.block0 = IFBlock(7 + 4, c=90)
+        self.block1 = IFBlock(7 + 4, c=90)
+        self.block2 = IFBlock(7 + 4, c=90)
+        self.block_tea = IFBlock(10 + 4, c=90)
+        # self.contextnet = Contextnet()
+        # self.unet = Unet()
+
+    def forward(self, x, scale_list=[4, 2, 1], training=False):
+        if training == False:
+            channel = x.shape[1] // 2
+            img0 = x[:, :channel]
+            img1 = x[:, channel:]
+        flow_list = []
+        merged = []
+        mask_list = []
+        warped_img0 = img0
+        warped_img1 = img1
+        flow = (x[:, :4]).detach() * 0
+        mask = (x[:, :1]).detach() * 0
+        loss_cons = 0
+        block = [self.block0, self.block1, self.block2]
+        for i in range(3):
+            f0, m0 = block[i](torch.cat((warped_img0[:, :3], warped_img1[:, :3], mask), 1), flow, scale=scale_list[i])
+            f1, m1 = block[i](
+                torch.cat((warped_img1[:, :3], warped_img0[:, :3], -mask), 1),
+                torch.cat((flow[:, 2:4], flow[:, :2]), 1),
+                scale=scale_list[i],
+            )
+            flow = flow + (f0 + torch.cat((f1[:, 2:4], f1[:, :2]), 1)) / 2
+            mask = mask + (m0 + (-m1)) / 2
+            mask_list.append(mask)
+            flow_list.append(flow)
+            warped_img0 = warp(img0, flow[:, :2])
+            warped_img1 = warp(img1, flow[:, 2:4])
+            merged.append((warped_img0, warped_img1))
+        """
+        c0 = self.contextnet(img0, flow[:, :2])
+        c1 = self.contextnet(img1, flow[:, 2:4])
+        tmp = self.unet(img0, img1, warped_img0, warped_img1, mask, flow, c0, c1)
+        res = tmp[:, 1:4] * 2 - 1
+        """
+        for i in range(3):
+            mask_list[i] = torch.sigmoid(mask_list[i])
+            merged[i] = merged[i][0] * mask_list[i] + merged[i][1] * (1 - mask_list[i])
+            # merged[i] = torch.clamp(merged[i] + res, 0, 1)
+        return flow_list, mask_list[2], merged

inference/gradio_composite_demo/rife/IFNet_m.py

@@ -0,0 +1,127 @@
+from .refine import *
+
+
+def deconv(in_planes, out_planes, kernel_size=4, stride=2, padding=1):
+    return nn.Sequential(
+        torch.nn.ConvTranspose2d(in_channels=in_planes, out_channels=out_planes, kernel_size=4, stride=2, padding=1),
+        nn.PReLU(out_planes),
+    )
+
+
+def conv(in_planes, out_planes, kernel_size=3, stride=1, padding=1, dilation=1):
+    return nn.Sequential(
+        nn.Conv2d(
+            in_planes,
+            out_planes,
+            kernel_size=kernel_size,
+            stride=stride,
+            padding=padding,
+            dilation=dilation,
+            bias=True,
+        ),
+        nn.PReLU(out_planes),
+    )
+
+
+class IFBlock(nn.Module):
+    def __init__(self, in_planes, c=64):
+        super(IFBlock, self).__init__()
+        self.conv0 = nn.Sequential(
+            conv(in_planes, c // 2, 3, 2, 1),
+            conv(c // 2, c, 3, 2, 1),
+        )
+        self.convblock = nn.Sequential(
+            conv(c, c),
+            conv(c, c),
+            conv(c, c),
+            conv(c, c),
+            conv(c, c),
+            conv(c, c),
+            conv(c, c),
+            conv(c, c),
+        )
+        self.lastconv = nn.ConvTranspose2d(c, 5, 4, 2, 1)
+
+    def forward(self, x, flow, scale):
+        if scale != 1:
+            x = F.interpolate(x, scale_factor=1.0 / scale, mode="bilinear", align_corners=False)
+        if flow != None:
+            flow = F.interpolate(flow, scale_factor=1.0 / scale, mode="bilinear", align_corners=False) * 1.0 / scale
+            x = torch.cat((x, flow), 1)
+        x = self.conv0(x)
+        x = self.convblock(x) + x
+        tmp = self.lastconv(x)
+        tmp = F.interpolate(tmp, scale_factor=scale * 2, mode="bilinear", align_corners=False)
+        flow = tmp[:, :4] * scale * 2
+        mask = tmp[:, 4:5]
+        return flow, mask
+
+
+class IFNet_m(nn.Module):
+    def __init__(self):
+        super(IFNet_m, self).__init__()
+        self.block0 = IFBlock(6 + 1, c=240)
+        self.block1 = IFBlock(13 + 4 + 1, c=150)
+        self.block2 = IFBlock(13 + 4 + 1, c=90)
+        self.block_tea = IFBlock(16 + 4 + 1, c=90)
+        self.contextnet = Contextnet()
+        self.unet = Unet()
+
+    def forward(self, x, scale=[4, 2, 1], timestep=0.5, returnflow=False):
+        timestep = (x[:, :1].clone() * 0 + 1) * timestep
+        img0 = x[:, :3]
+        img1 = x[:, 3:6]
+        gt = x[:, 6:]  # In inference time, gt is None
+        flow_list = []
+        merged = []
+        mask_list = []
+        warped_img0 = img0
+        warped_img1 = img1
+        flow = None
+        loss_distill = 0
+        stu = [self.block0, self.block1, self.block2]
+        for i in range(3):
+            if flow != None:
+                flow_d, mask_d = stu[i](
+                    torch.cat((img0, img1, timestep, warped_img0, warped_img1, mask), 1), flow, scale=scale[i]
+                )
+                flow = flow + flow_d
+                mask = mask + mask_d
+            else:
+                flow, mask = stu[i](torch.cat((img0, img1, timestep), 1), None, scale=scale[i])
+            mask_list.append(torch.sigmoid(mask))
+            flow_list.append(flow)
+            warped_img0 = warp(img0, flow[:, :2])
+            warped_img1 = warp(img1, flow[:, 2:4])
+            merged_student = (warped_img0, warped_img1)
+            merged.append(merged_student)
+        if gt.shape[1] == 3:
+            flow_d, mask_d = self.block_tea(
+                torch.cat((img0, img1, timestep, warped_img0, warped_img1, mask, gt), 1), flow, scale=1
+            )
+            flow_teacher = flow + flow_d
+            warped_img0_teacher = warp(img0, flow_teacher[:, :2])
+            warped_img1_teacher = warp(img1, flow_teacher[:, 2:4])
+            mask_teacher = torch.sigmoid(mask + mask_d)
+            merged_teacher = warped_img0_teacher * mask_teacher + warped_img1_teacher * (1 - mask_teacher)
+        else:
+            flow_teacher = None
+            merged_teacher = None
+        for i in range(3):
+            merged[i] = merged[i][0] * mask_list[i] + merged[i][1] * (1 - mask_list[i])
+            if gt.shape[1] == 3:
+                loss_mask = (
+                    ((merged[i] - gt).abs().mean(1, True) > (merged_teacher - gt).abs().mean(1, True) + 0.01)
+                    .float()
+                    .detach()
+                )
+                loss_distill += (((flow_teacher.detach() - flow_list[i]) ** 2).mean(1, True) ** 0.5 * loss_mask).mean()
+        if returnflow:
+            return flow
+        else:
+            c0 = self.contextnet(img0, flow[:, :2])
+            c1 = self.contextnet(img1, flow[:, 2:4])
+            tmp = self.unet(img0, img1, warped_img0, warped_img1, mask, flow, c0, c1)
+            res = tmp[:, :3] * 2 - 1
+            merged[2] = torch.clamp(merged[2] + res, 0, 1)
+        return flow_list, mask_list[2], merged, flow_teacher, merged_teacher, loss_distill

inference/gradio_composite_demo/rife/RIFE.py

@@ -0,0 +1,95 @@
+from torch.optim import AdamW
+from torch.nn.parallel import DistributedDataParallel as DDP
+from .IFNet import *
+from .IFNet_m import *
+from .loss import *
+from .laplacian import *
+from .refine import *
+
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+
+class Model:
+    def __init__(self, local_rank=-1, arbitrary=False):
+        if arbitrary == True:
+            self.flownet = IFNet_m()
+        else:
+            self.flownet = IFNet()
+        self.device()
+        self.optimG = AdamW(
+            self.flownet.parameters(), lr=1e-6, weight_decay=1e-3
+        )  # use large weight decay may avoid NaN loss
+        self.epe = EPE()
+        self.lap = LapLoss()
+        self.sobel = SOBEL()
+        if local_rank != -1:
+            self.flownet = DDP(self.flownet, device_ids=[local_rank], output_device=local_rank)
+
+    def train(self):
+        self.flownet.train()
+
+    def eval(self):
+        self.flownet.eval()
+
+    def device(self):
+        self.flownet.to(device)
+
+    def load_model(self, path, rank=0):
+        def convert(param):
+            return {k.replace("module.", ""): v for k, v in param.items() if "module." in k}
+
+        if rank <= 0:
+            self.flownet.load_state_dict(convert(torch.load("{}/flownet.pkl".format(path))))
+
+    def save_model(self, path, rank=0):
+        if rank == 0:
+            torch.save(self.flownet.state_dict(), "{}/flownet.pkl".format(path))
+
+    def inference(self, img0, img1, scale=1, scale_list=[4, 2, 1], TTA=False, timestep=0.5):
+        for i in range(3):
+            scale_list[i] = scale_list[i] * 1.0 / scale
+        imgs = torch.cat((img0, img1), 1)
+        flow, mask, merged, flow_teacher, merged_teacher, loss_distill = self.flownet(
+            imgs, scale_list, timestep=timestep
+        )
+        if TTA == False:
+            return merged[2]
+        else:
+            flow2, mask2, merged2, flow_teacher2, merged_teacher2, loss_distill2 = self.flownet(
+                imgs.flip(2).flip(3), scale_list, timestep=timestep
+            )
+            return (merged[2] + merged2[2].flip(2).flip(3)) / 2
+
+    def update(self, imgs, gt, learning_rate=0, mul=1, training=True, flow_gt=None):
+        for param_group in self.optimG.param_groups:
+            param_group["lr"] = learning_rate
+        img0 = imgs[:, :3]
+        img1 = imgs[:, 3:]
+        if training:
+            self.train()
+        else:
+            self.eval()
+        flow, mask, merged, flow_teacher, merged_teacher, loss_distill = self.flownet(
+            torch.cat((imgs, gt), 1), scale=[4, 2, 1]
+        )
+        loss_l1 = (self.lap(merged[2], gt)).mean()
+        loss_tea = (self.lap(merged_teacher, gt)).mean()
+        if training:
+            self.optimG.zero_grad()
+            loss_G = (
+                loss_l1 + loss_tea + loss_distill * 0.01
+            )  # when training RIFEm, the weight of loss_distill should be 0.005 or 0.002
+            loss_G.backward()
+            self.optimG.step()
+        else:
+            flow_teacher = flow[2]
+        return merged[2], {
+            "merged_tea": merged_teacher,
+            "mask": mask,
+            "mask_tea": mask,
+            "flow": flow[2][:, :2],
+            "flow_tea": flow_teacher,
+            "loss_l1": loss_l1,
+            "loss_tea": loss_tea,
+            "loss_distill": loss_distill,
+        }

inference/gradio_composite_demo/rife/RIFE_HDv3.py

@@ -0,0 +1,86 @@
+import torch
+import torch.nn as nn
+import numpy as np
+from torch.optim import AdamW
+import torch.optim as optim
+import itertools
+from .warplayer import warp
+from torch.nn.parallel import DistributedDataParallel as DDP
+from .IFNet_HDv3 import *
+import torch.nn.functional as F
+from .loss import *
+
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+
+class Model:
+    def __init__(self, local_rank=-1):
+        self.flownet = IFNet()
+        self.device()
+        self.optimG = AdamW(self.flownet.parameters(), lr=1e-6, weight_decay=1e-4)
+        self.epe = EPE()
+        # self.vgg = VGGPerceptualLoss().to(device)
+        self.sobel = SOBEL()
+        if local_rank != -1:
+            self.flownet = DDP(self.flownet, device_ids=[local_rank], output_device=local_rank)
+
+    def train(self):
+        self.flownet.train()
+
+    def eval(self):
+        self.flownet.eval()
+
+    def device(self):
+        self.flownet.to(device)
+
+    def load_model(self, path, rank=0):
+        def convert(param):
+            if rank == -1:
+                return {k.replace("module.", ""): v for k, v in param.items() if "module." in k}
+            else:
+                return param
+
+        if rank <= 0:
+            if torch.cuda.is_available():
+                self.flownet.load_state_dict(convert(torch.load("{}/flownet.pkl".format(path))))
+            else:
+                self.flownet.load_state_dict(convert(torch.load("{}/flownet.pkl".format(path), map_location="cpu")))
+
+    def save_model(self, path, rank=0):
+        if rank == 0:
+            torch.save(self.flownet.state_dict(), "{}/flownet.pkl".format(path))
+
+    def inference(self, img0, img1, scale=1.0):
+        imgs = torch.cat((img0, img1), 1)
+        scale_list = [4 / scale, 2 / scale, 1 / scale]
+        flow, mask, merged = self.flownet(imgs, scale_list)
+        return merged[2]
+
+    def update(self, imgs, gt, learning_rate=0, mul=1, training=True, flow_gt=None):
+        for param_group in self.optimG.param_groups:
+            param_group["lr"] = learning_rate
+        img0 = imgs[:, :3]
+        img1 = imgs[:, 3:]
+        if training:
+            self.train()
+        else:
+            self.eval()
+        scale = [4, 2, 1]
+        flow, mask, merged = self.flownet(torch.cat((imgs, gt), 1), scale=scale, training=training)
+        loss_l1 = (merged[2] - gt).abs().mean()
+        loss_smooth = self.sobel(flow[2], flow[2] * 0).mean()
+        # loss_vgg = self.vgg(merged[2], gt)
+        if training:
+            self.optimG.zero_grad()
+            loss_G = loss_cons + loss_smooth * 0.1
+            loss_G.backward()
+            self.optimG.step()
+        else:
+            flow_teacher = flow[2]
+        return merged[2], {
+            "mask": mask,
+            "flow": flow[2][:, :2],
+            "loss_l1": loss_l1,
+            "loss_cons": loss_cons,
+            "loss_smooth": loss_smooth,
+        }

inference/gradio_composite_demo/rife/laplacian.py

@@ -0,0 +1,69 @@
+import torch
+import numpy as np
+import torch.nn as nn
+import torch.nn.functional as F
+
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+import torch
+
+
+def gauss_kernel(size=5, channels=3):
+    kernel = torch.tensor(
+        [
+            [1.0, 4.0, 6.0, 4.0, 1],
+            [4.0, 16.0, 24.0, 16.0, 4.0],
+            [6.0, 24.0, 36.0, 24.0, 6.0],
+            [4.0, 16.0, 24.0, 16.0, 4.0],
+            [1.0, 4.0, 6.0, 4.0, 1.0],
+        ]
+    )
+    kernel /= 256.0
+    kernel = kernel.repeat(channels, 1, 1, 1)
+    kernel = kernel.to(device)
+    return kernel
+
+
+def downsample(x):
+    return x[:, :, ::2, ::2]
+
+
+def upsample(x):
+    cc = torch.cat([x, torch.zeros(x.shape[0], x.shape[1], x.shape[2], x.shape[3]).to(device)], dim=3)
+    cc = cc.view(x.shape[0], x.shape[1], x.shape[2] * 2, x.shape[3])
+    cc = cc.permute(0, 1, 3, 2)
+    cc = torch.cat([cc, torch.zeros(x.shape[0], x.shape[1], x.shape[3], x.shape[2] * 2).to(device)], dim=3)
+    cc = cc.view(x.shape[0], x.shape[1], x.shape[3] * 2, x.shape[2] * 2)
+    x_up = cc.permute(0, 1, 3, 2)
+    return conv_gauss(x_up, 4 * gauss_kernel(channels=x.shape[1]))
+
+
+def conv_gauss(img, kernel):
+    img = torch.nn.functional.pad(img, (2, 2, 2, 2), mode="reflect")
+    out = torch.nn.functional.conv2d(img, kernel, groups=img.shape[1])
+    return out
+
+
+def laplacian_pyramid(img, kernel, max_levels=3):
+    current = img
+    pyr = []
+    for level in range(max_levels):
+        filtered = conv_gauss(current, kernel)
+        down = downsample(filtered)
+        up = upsample(down)
+        diff = current - up
+        pyr.append(diff)
+        current = down
+    return pyr
+
+
+class LapLoss(torch.nn.Module):
+    def __init__(self, max_levels=5, channels=3):
+        super(LapLoss, self).__init__()
+        self.max_levels = max_levels
+        self.gauss_kernel = gauss_kernel(channels=channels)
+
+    def forward(self, input, target):
+        pyr_input = laplacian_pyramid(img=input, kernel=self.gauss_kernel, max_levels=self.max_levels)
+        pyr_target = laplacian_pyramid(img=target, kernel=self.gauss_kernel, max_levels=self.max_levels)
+        return sum(torch.nn.functional.l1_loss(a, b) for a, b in zip(pyr_input, pyr_target))

inference/gradio_composite_demo/rife/loss.py

@@ -0,0 +1,130 @@
+import torch
+import numpy as np
+import torch.nn as nn
+import torch.nn.functional as F
+import torchvision.models as models
+
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+
+class EPE(nn.Module):
+    def __init__(self):
+        super(EPE, self).__init__()
+
+    def forward(self, flow, gt, loss_mask):
+        loss_map = (flow - gt.detach()) ** 2
+        loss_map = (loss_map.sum(1, True) + 1e-6) ** 0.5
+        return loss_map * loss_mask
+
+
+class Ternary(nn.Module):
+    def __init__(self):
+        super(Ternary, self).__init__()
+        patch_size = 7
+        out_channels = patch_size * patch_size
+        self.w = np.eye(out_channels).reshape((patch_size, patch_size, 1, out_channels))
+        self.w = np.transpose(self.w, (3, 2, 0, 1))
+        self.w = torch.tensor(self.w).float().to(device)
+
+    def transform(self, img):
+        patches = F.conv2d(img, self.w, padding=3, bias=None)
+        transf = patches - img
+        transf_norm = transf / torch.sqrt(0.81 + transf**2)
+        return transf_norm
+
+    def rgb2gray(self, rgb):
+        r, g, b = rgb[:, 0:1, :, :], rgb[:, 1:2, :, :], rgb[:, 2:3, :, :]
+        gray = 0.2989 * r + 0.5870 * g + 0.1140 * b
+        return gray
+
+    def hamming(self, t1, t2):
+        dist = (t1 - t2) ** 2
+        dist_norm = torch.mean(dist / (0.1 + dist), 1, True)
+        return dist_norm
+
+    def valid_mask(self, t, padding):
+        n, _, h, w = t.size()
+        inner = torch.ones(n, 1, h - 2 * padding, w - 2 * padding).type_as(t)
+        mask = F.pad(inner, [padding] * 4)
+        return mask
+
+    def forward(self, img0, img1):
+        img0 = self.transform(self.rgb2gray(img0))
+        img1 = self.transform(self.rgb2gray(img1))
+        return self.hamming(img0, img1) * self.valid_mask(img0, 1)
+
+
+class SOBEL(nn.Module):
+    def __init__(self):
+        super(SOBEL, self).__init__()
+        self.kernelX = torch.tensor(
+            [
+                [1, 0, -1],
+                [2, 0, -2],
+                [1, 0, -1],
+            ]
+        ).float()
+        self.kernelY = self.kernelX.clone().T
+        self.kernelX = self.kernelX.unsqueeze(0).unsqueeze(0).to(device)
+        self.kernelY = self.kernelY.unsqueeze(0).unsqueeze(0).to(device)
+
+    def forward(self, pred, gt):
+        N, C, H, W = pred.shape[0], pred.shape[1], pred.shape[2], pred.shape[3]
+        img_stack = torch.cat([pred.reshape(N * C, 1, H, W), gt.reshape(N * C, 1, H, W)], 0)
+        sobel_stack_x = F.conv2d(img_stack, self.kernelX, padding=1)
+        sobel_stack_y = F.conv2d(img_stack, self.kernelY, padding=1)
+        pred_X, gt_X = sobel_stack_x[: N * C], sobel_stack_x[N * C :]
+        pred_Y, gt_Y = sobel_stack_y[: N * C], sobel_stack_y[N * C :]
+
+        L1X, L1Y = torch.abs(pred_X - gt_X), torch.abs(pred_Y - gt_Y)
+        loss = L1X + L1Y
+        return loss
+
+
+class MeanShift(nn.Conv2d):
+    def __init__(self, data_mean, data_std, data_range=1, norm=True):
+        c = len(data_mean)
+        super(MeanShift, self).__init__(c, c, kernel_size=1)
+        std = torch.Tensor(data_std)
+        self.weight.data = torch.eye(c).view(c, c, 1, 1)
+        if norm:
+            self.weight.data.div_(std.view(c, 1, 1, 1))
+            self.bias.data = -1 * data_range * torch.Tensor(data_mean)
+            self.bias.data.div_(std)
+        else:
+            self.weight.data.mul_(std.view(c, 1, 1, 1))
+            self.bias.data = data_range * torch.Tensor(data_mean)
+        self.requires_grad = False
+
+
+class VGGPerceptualLoss(torch.nn.Module):
+    def __init__(self, rank=0):
+        super(VGGPerceptualLoss, self).__init__()
+        blocks = []
+        pretrained = True
+        self.vgg_pretrained_features = models.vgg19(pretrained=pretrained).features
+        self.normalize = MeanShift([0.485, 0.456, 0.406], [0.229, 0.224, 0.225], norm=True).cuda()
+        for param in self.parameters():
+            param.requires_grad = False
+
+    def forward(self, X, Y, indices=None):
+        X = self.normalize(X)
+        Y = self.normalize(Y)
+        indices = [2, 7, 12, 21, 30]
+        weights = [1.0 / 2.6, 1.0 / 4.8, 1.0 / 3.7, 1.0 / 5.6, 10 / 1.5]
+        k = 0
+        loss = 0
+        for i in range(indices[-1]):
+            X = self.vgg_pretrained_features[i](X)
+            Y = self.vgg_pretrained_features[i](Y)
+            if (i + 1) in indices:
+                loss += weights[k] * (X - Y.detach()).abs().mean() * 0.1
+                k += 1
+        return loss
+
+
+if __name__ == "__main__":
+    img0 = torch.zeros(3, 3, 256, 256).float().to(device)
+    img1 = torch.tensor(np.random.normal(0, 1, (3, 3, 256, 256))).float().to(device)
+    ternary_loss = Ternary()
+    print(ternary_loss(img0, img1).shape)

inference/gradio_composite_demo/rife/pytorch_msssim/__init__.py

@@ -0,0 +1,203 @@
+import torch
+import torch.nn.functional as F
+from math import exp
+import numpy as np
+
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+
+def gaussian(window_size, sigma):
+    gauss = torch.Tensor([exp(-((x - window_size // 2) ** 2) / float(2 * sigma**2)) for x in range(window_size)])
+    return gauss / gauss.sum()
+
+
+def create_window(window_size, channel=1):
+    _1D_window = gaussian(window_size, 1.5).unsqueeze(1)
+    _2D_window = _1D_window.mm(_1D_window.t()).float().unsqueeze(0).unsqueeze(0).to(device)
+    window = _2D_window.expand(channel, 1, window_size, window_size).contiguous()
+    return window
+
+
+def create_window_3d(window_size, channel=1):
+    _1D_window = gaussian(window_size, 1.5).unsqueeze(1)
+    _2D_window = _1D_window.mm(_1D_window.t())
+    _3D_window = _2D_window.unsqueeze(2) @ (_1D_window.t())
+    window = _3D_window.expand(1, channel, window_size, window_size, window_size).contiguous().to(device)
+    return window
+
+
+def ssim(img1, img2, window_size=11, window=None, size_average=True, full=False, val_range=None):
+    # Value range can be different from 255. Other common ranges are 1 (sigmoid) and 2 (tanh).
+    if val_range is None:
+        if torch.max(img1) > 128:
+            max_val = 255
+        else:
+            max_val = 1
+
+        if torch.min(img1) < -0.5:
+            min_val = -1
+        else:
+            min_val = 0
+        L = max_val - min_val
+    else:
+        L = val_range
+
+    padd = 0
+    (_, channel, height, width) = img1.size()
+    if window is None:
+        real_size = min(window_size, height, width)
+        window = create_window(real_size, channel=channel).to(img1.device)
+
+    # mu1 = F.conv2d(img1, window, padding=padd, groups=channel)
+    # mu2 = F.conv2d(img2, window, padding=padd, groups=channel)
+    mu1 = F.conv2d(F.pad(img1, (5, 5, 5, 5), mode="replicate"), window, padding=padd, groups=channel)
+    mu2 = F.conv2d(F.pad(img2, (5, 5, 5, 5), mode="replicate"), window, padding=padd, groups=channel)
+
+    mu1_sq = mu1.pow(2)
+    mu2_sq = mu2.pow(2)
+    mu1_mu2 = mu1 * mu2
+
+    sigma1_sq = F.conv2d(F.pad(img1 * img1, (5, 5, 5, 5), "replicate"), window, padding=padd, groups=channel) - mu1_sq
+    sigma2_sq = F.conv2d(F.pad(img2 * img2, (5, 5, 5, 5), "replicate"), window, padding=padd, groups=channel) - mu2_sq
+    sigma12 = F.conv2d(F.pad(img1 * img2, (5, 5, 5, 5), "replicate"), window, padding=padd, groups=channel) - mu1_mu2
+
+    C1 = (0.01 * L) ** 2
+    C2 = (0.03 * L) ** 2
+
+    v1 = 2.0 * sigma12 + C2
+    v2 = sigma1_sq + sigma2_sq + C2
+    cs = torch.mean(v1 / v2)  # contrast sensitivity
+
+    ssim_map = ((2 * mu1_mu2 + C1) * v1) / ((mu1_sq + mu2_sq + C1) * v2)
+
+    if size_average:
+        ret = ssim_map.mean()
+    else:
+        ret = ssim_map.mean(1).mean(1).mean(1)
+
+    if full:
+        return ret, cs
+    return ret
+
+
+def ssim_matlab(img1, img2, window_size=11, window=None, size_average=True, full=False, val_range=None):
+    # Value range can be different from 255. Other common ranges are 1 (sigmoid) and 2 (tanh).
+    if val_range is None:
+        if torch.max(img1) > 128:
+            max_val = 255
+        else:
+            max_val = 1
+
+        if torch.min(img1) < -0.5:
+            min_val = -1
+        else:
+            min_val = 0
+        L = max_val - min_val
+    else:
+        L = val_range
+
+    padd = 0
+    (_, _, height, width) = img1.size()
+    if window is None:
+        real_size = min(window_size, height, width)
+        window = create_window_3d(real_size, channel=1).to(img1.device, dtype=img1.dtype)
+        # Channel is set to 1 since we consider color images as volumetric images
+
+    img1 = img1.unsqueeze(1)
+    img2 = img2.unsqueeze(1)
+
+    mu1 = F.conv3d(F.pad(img1, (5, 5, 5, 5, 5, 5), mode="replicate"), window, padding=padd, groups=1)
+    mu2 = F.conv3d(F.pad(img2, (5, 5, 5, 5, 5, 5), mode="replicate"), window, padding=padd, groups=1)
+
+    mu1_sq = mu1.pow(2)
+    mu2_sq = mu2.pow(2)
+    mu1_mu2 = mu1 * mu2
+
+    sigma1_sq = F.conv3d(F.pad(img1 * img1, (5, 5, 5, 5, 5, 5), "replicate"), window, padding=padd, groups=1) - mu1_sq
+    sigma2_sq = F.conv3d(F.pad(img2 * img2, (5, 5, 5, 5, 5, 5), "replicate"), window, padding=padd, groups=1) - mu2_sq
+    sigma12 = F.conv3d(F.pad(img1 * img2, (5, 5, 5, 5, 5, 5), "replicate"), window, padding=padd, groups=1) - mu1_mu2
+
+    C1 = (0.01 * L) ** 2
+    C2 = (0.03 * L) ** 2
+
+    v1 = 2.0 * sigma12 + C2
+    v2 = sigma1_sq + sigma2_sq + C2
+    cs = torch.mean(v1 / v2)  # contrast sensitivity
+
+    ssim_map = ((2 * mu1_mu2 + C1) * v1) / ((mu1_sq + mu2_sq + C1) * v2)
+
+    if size_average:
+        ret = ssim_map.mean()
+    else:
+        ret = ssim_map.mean(1).mean(1).mean(1)
+
+    if full:
+        return ret, cs
+    return ret
+
+
+def msssim(img1, img2, window_size=11, size_average=True, val_range=None, normalize=False):
+    device = img1.device
+    weights = torch.FloatTensor([0.0448, 0.2856, 0.3001, 0.2363, 0.1333]).to(device)
+    levels = weights.size()[0]
+    mssim = []
+    mcs = []
+    for _ in range(levels):
+        sim, cs = ssim(img1, img2, window_size=window_size, size_average=size_average, full=True, val_range=val_range)
+        mssim.append(sim)
+        mcs.append(cs)
+
+        img1 = F.avg_pool2d(img1, (2, 2))
+        img2 = F.avg_pool2d(img2, (2, 2))
+
+    mssim = torch.stack(mssim)
+    mcs = torch.stack(mcs)
+
+    # Normalize (to avoid NaNs during training unstable models, not compliant with original definition)
+    if normalize:
+        mssim = (mssim + 1) / 2
+        mcs = (mcs + 1) / 2
+
+    pow1 = mcs**weights
+    pow2 = mssim**weights
+    # From Matlab implementation https://ece.uwaterloo.ca/~z70wang/research/iwssim/
+    output = torch.prod(pow1[:-1] * pow2[-1])
+    return output
+
+
+# Classes to re-use window
+class SSIM(torch.nn.Module):
+    def __init__(self, window_size=11, size_average=True, val_range=None):
+        super(SSIM, self).__init__()
+        self.window_size = window_size
+        self.size_average = size_average
+        self.val_range = val_range
+
+        # Assume 3 channel for SSIM
+        self.channel = 3
+        self.window = create_window(window_size, channel=self.channel)
+
+    def forward(self, img1, img2):
+        (_, channel, _, _) = img1.size()
+
+        if channel == self.channel and self.window.dtype == img1.dtype:
+            window = self.window
+        else:
+            window = create_window(self.window_size, channel).to(img1.device).type(img1.dtype)
+            self.window = window
+            self.channel = channel
+
+        _ssim = ssim(img1, img2, window=window, window_size=self.window_size, size_average=self.size_average)
+        dssim = (1 - _ssim) / 2
+        return dssim
+
+
+class MSSSIM(torch.nn.Module):
+    def __init__(self, window_size=11, size_average=True, channel=3):
+        super(MSSSIM, self).__init__()
+        self.window_size = window_size
+        self.size_average = size_average
+        self.channel = channel
+
+    def forward(self, img1, img2):
+        return msssim(img1, img2, window_size=self.window_size, size_average=self.size_average)

inference/gradio_composite_demo/rife/refine.py

@@ -0,0 +1,107 @@
+import torch
+import torch.nn as nn
+from .warplayer import warp
+import torch.nn.functional as F
+
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+
+def conv(in_planes, out_planes, kernel_size=3, stride=1, padding=1, dilation=1):
+    return nn.Sequential(
+        nn.Conv2d(
+            in_planes,
+            out_planes,
+            kernel_size=kernel_size,
+            stride=stride,
+            padding=padding,
+            dilation=dilation,
+            bias=True,
+        ),
+        nn.PReLU(out_planes),
+    )
+
+
+def deconv(in_planes, out_planes, kernel_size=4, stride=2, padding=1):
+    return nn.Sequential(
+        torch.nn.ConvTranspose2d(
+            in_channels=in_planes, out_channels=out_planes, kernel_size=4, stride=2, padding=1, bias=True
+        ),
+        nn.PReLU(out_planes),
+    )
+
+
+class Conv2(nn.Module):
+    def __init__(self, in_planes, out_planes, stride=2):
+        super(Conv2, self).__init__()
+        self.conv1 = conv(in_planes, out_planes, 3, stride, 1)
+        self.conv2 = conv(out_planes, out_planes, 3, 1, 1)
+
+    def forward(self, x):
+        x = self.conv1(x)
+        x = self.conv2(x)
+        return x
+
+
+c = 16
+
+
+class Contextnet(nn.Module):
+    def __init__(self):
+        super(Contextnet, self).__init__()
+        self.conv1 = Conv2(3, c)
+        self.conv2 = Conv2(c, 2 * c)
+        self.conv3 = Conv2(2 * c, 4 * c)
+        self.conv4 = Conv2(4 * c, 8 * c)
+
+    def forward(self, x, flow):
+        x = self.conv1(x)
+        flow = (
+            F.interpolate(flow, scale_factor=0.5, mode="bilinear", align_corners=False, recompute_scale_factor=False)
+            * 0.5
+        )
+        f1 = warp(x, flow)
+        x = self.conv2(x)
+        flow = (
+            F.interpolate(flow, scale_factor=0.5, mode="bilinear", align_corners=False, recompute_scale_factor=False)
+            * 0.5
+        )
+        f2 = warp(x, flow)
+        x = self.conv3(x)
+        flow = (
+            F.interpolate(flow, scale_factor=0.5, mode="bilinear", align_corners=False, recompute_scale_factor=False)
+            * 0.5
+        )
+        f3 = warp(x, flow)
+        x = self.conv4(x)
+        flow = (
+            F.interpolate(flow, scale_factor=0.5, mode="bilinear", align_corners=False, recompute_scale_factor=False)
+            * 0.5
+        )
+        f4 = warp(x, flow)
+        return [f1, f2, f3, f4]
+
+
+class Unet(nn.Module):
+    def __init__(self):
+        super(Unet, self).__init__()
+        self.down0 = Conv2(17, 2 * c)
+        self.down1 = Conv2(4 * c, 4 * c)
+        self.down2 = Conv2(8 * c, 8 * c)
+        self.down3 = Conv2(16 * c, 16 * c)
+        self.up0 = deconv(32 * c, 8 * c)
+        self.up1 = deconv(16 * c, 4 * c)
+        self.up2 = deconv(8 * c, 2 * c)
+        self.up3 = deconv(4 * c, c)
+        self.conv = nn.Conv2d(c, 3, 3, 1, 1)
+
+    def forward(self, img0, img1, warped_img0, warped_img1, mask, flow, c0, c1):
+        s0 = self.down0(torch.cat((img0, img1, warped_img0, warped_img1, mask, flow), 1))
+        s1 = self.down1(torch.cat((s0, c0[0], c1[0]), 1))
+        s2 = self.down2(torch.cat((s1, c0[1], c1[1]), 1))
+        s3 = self.down3(torch.cat((s2, c0[2], c1[2]), 1))
+        x = self.up0(torch.cat((s3, c0[3], c1[3]), 1))
+        x = self.up1(torch.cat((x, s2), 1))
+        x = self.up2(torch.cat((x, s1), 1))
+        x = self.up3(torch.cat((x, s0), 1))
+        x = self.conv(x)
+        return torch.sigmoid(x)

inference/gradio_composite_demo/rife/refine_2R.py

@@ -0,0 +1,104 @@
+import torch
+import torch.nn as nn
+from .warplayer import warp
+import torch.nn.functional as F
+
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+
+def conv(in_planes, out_planes, kernel_size=3, stride=1, padding=1, dilation=1):
+    return nn.Sequential(
+        nn.Conv2d(
+            in_planes,
+            out_planes,
+            kernel_size=kernel_size,
+            stride=stride,
+            padding=padding,
+            dilation=dilation,
+            bias=True,
+        ),
+        nn.PReLU(out_planes),
+    )
+
+
+def deconv(in_planes, out_planes, kernel_size=4, stride=2, padding=1):
+    return nn.Sequential(
+        torch.nn.ConvTranspose2d(
+            in_channels=in_planes, out_channels=out_planes, kernel_size=4, stride=2, padding=1, bias=True
+        ),
+        nn.PReLU(out_planes),
+    )
+
+
+class Conv2(nn.Module):
+    def __init__(self, in_planes, out_planes, stride=2):
+        super(Conv2, self).__init__()
+        self.conv1 = conv(in_planes, out_planes, 3, stride, 1)
+        self.conv2 = conv(out_planes, out_planes, 3, 1, 1)
+
+    def forward(self, x):
+        x = self.conv1(x)
+        x = self.conv2(x)
+        return x
+
+
+c = 16
+
+
+class Contextnet(nn.Module):
+    def __init__(self):
+        super(Contextnet, self).__init__()
+        self.conv1 = Conv2(3, c, 1)
+        self.conv2 = Conv2(c, 2 * c)
+        self.conv3 = Conv2(2 * c, 4 * c)
+        self.conv4 = Conv2(4 * c, 8 * c)
+
+    def forward(self, x, flow):
+        x = self.conv1(x)
+        # flow = F.interpolate(flow, scale_factor=0.5, mode="bilinear", align_corners=False, recompute_scale_factor=False) * 0.5
+        f1 = warp(x, flow)
+        x = self.conv2(x)
+        flow = (
+            F.interpolate(flow, scale_factor=0.5, mode="bilinear", align_corners=False, recompute_scale_factor=False)
+            * 0.5
+        )
+        f2 = warp(x, flow)
+        x = self.conv3(x)
+        flow = (
+            F.interpolate(flow, scale_factor=0.5, mode="bilinear", align_corners=False, recompute_scale_factor=False)
+            * 0.5
+        )
+        f3 = warp(x, flow)
+        x = self.conv4(x)
+        flow = (
+            F.interpolate(flow, scale_factor=0.5, mode="bilinear", align_corners=False, recompute_scale_factor=False)
+            * 0.5
+        )
+        f4 = warp(x, flow)
+        return [f1, f2, f3, f4]
+
+
+class Unet(nn.Module):
+    def __init__(self):
+        super(Unet, self).__init__()
+        self.down0 = Conv2(17, 2 * c, 1)
+        self.down1 = Conv2(4 * c, 4 * c)
+        self.down2 = Conv2(8 * c, 8 * c)
+        self.down3 = Conv2(16 * c, 16 * c)
+        self.up0 = deconv(32 * c, 8 * c)
+        self.up1 = deconv(16 * c, 4 * c)
+        self.up2 = deconv(8 * c, 2 * c)
+        self.up3 = deconv(4 * c, c)
+        self.conv = nn.Conv2d(c, 3, 3, 2, 1)
+
+    def forward(self, img0, img1, warped_img0, warped_img1, mask, flow, c0, c1):
+        s0 = self.down0(torch.cat((img0, img1, warped_img0, warped_img1, mask, flow), 1))
+        s1 = self.down1(torch.cat((s0, c0[0], c1[0]), 1))
+        s2 = self.down2(torch.cat((s1, c0[1], c1[1]), 1))
+        s3 = self.down3(torch.cat((s2, c0[2], c1[2]), 1))
+        x = self.up0(torch.cat((s3, c0[3], c1[3]), 1))
+        x = self.up1(torch.cat((x, s2), 1))
+        x = self.up2(torch.cat((x, s1), 1))
+        x = self.up3(torch.cat((x, s0), 1))
+        x = self.conv(x)
+        return torch.sigmoid(x)

inference/gradio_composite_demo/rife/warplayer.py

@@ -0,0 +1,34 @@
+import torch
+import torch.nn as nn
+
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+backwarp_tenGrid = {}
+
+
+def warp(tenInput, tenFlow):
+    k = (str(tenFlow.device), str(tenFlow.size()))
+    if k not in backwarp_tenGrid:
+        tenHorizontal = (
+            torch.linspace(-1.0, 1.0, tenFlow.shape[3], device=device)
+            .view(1, 1, 1, tenFlow.shape[3])
+            .expand(tenFlow.shape[0], -1, tenFlow.shape[2], -1)
+        )
+        tenVertical = (
+            torch.linspace(-1.0, 1.0, tenFlow.shape[2], device=device)
+            .view(1, 1, tenFlow.shape[2], 1)
+            .expand(tenFlow.shape[0], -1, -1, tenFlow.shape[3])
+        )
+        backwarp_tenGrid[k] = torch.cat([tenHorizontal, tenVertical], 1).to(device)
+
+    tenFlow = torch.cat(
+        [
+            tenFlow[:, 0:1, :, :] / ((tenInput.shape[3] - 1.0) / 2.0),
+            tenFlow[:, 1:2, :, :] / ((tenInput.shape[2] - 1.0) / 2.0),
+        ],
+        1,
+    )
+
+    g = (backwarp_tenGrid[k] + tenFlow).permute(0, 2, 3, 1)
+    return torch.nn.functional.grid_sample(
+        input=tenInput, grid=g, mode="bilinear", padding_mode="border", align_corners=True
+    )

inference/gradio_composite_demo/rife_model.py

@@ -0,0 +1,133 @@
+import torch
+from diffusers.image_processor import VaeImageProcessor
+from torch.nn import functional as F
+import cv2
+import utils
+from rife.pytorch_msssim import ssim_matlab
+import numpy as np
+import logging
+import skvideo.io
+from rife.RIFE_HDv3 import Model
+
+logger = logging.getLogger(__name__)
+
+device = "cuda" if torch.cuda.is_available() else "cpu"
+
+
+def pad_image(img, scale):
+    _, _, h, w = img.shape
+    tmp = max(32, int(32 / scale))
+    ph = ((h - 1) // tmp + 1) * tmp
+    pw = ((w - 1) // tmp + 1) * tmp
+    padding = (0, 0, pw - w, ph - h)
+    return F.pad(img, padding)
+
+
+def make_inference(model, I0, I1, upscale_amount, n):
+    middle = model.inference(I0, I1, upscale_amount)
+    if n == 1:
+        return [middle]
+    first_half = make_inference(model, I0, middle, upscale_amount, n=n // 2)
+    second_half = make_inference(model, middle, I1, upscale_amount, n=n // 2)
+    if n % 2:
+        return [*first_half, middle, *second_half]
+    else:
+        return [*first_half, *second_half]
+
+
+@torch.inference_mode()
+def ssim_interpolation_rife(model, samples, exp=1, upscale_amount=1, output_device="cpu"):
+    print(f"samples dtype:{samples.dtype}")
+    print(f"samples shape:{samples.shape}")
+    output = []
+    # [f, c, h, w]
+    for b in range(samples.shape[0]):
+        frame = samples[b : b + 1]
+        _, _, h, w = frame.shape
+        I0 = samples[b : b + 1]
+        I1 = samples[b + 1 : b + 2] if b + 2 < samples.shape[0] else samples[-1:]
+        I1 = pad_image(I1, upscale_amount)
+        # [c, h, w]
+        I0_small = F.interpolate(I0, (32, 32), mode="bilinear", align_corners=False)
+        I1_small = F.interpolate(I1, (32, 32), mode="bilinear", align_corners=False)
+
+        ssim = ssim_matlab(I0_small[:, :3], I1_small[:, :3])
+
+        if ssim > 0.996:
+            I1 = I0
+            I1 = pad_image(I1, upscale_amount)
+            I1 = make_inference(model, I0, I1, upscale_amount, 1)
+
+            I1_small = F.interpolate(I1[0], (32, 32), mode="bilinear", align_corners=False)
+            ssim = ssim_matlab(I0_small[:, :3], I1_small[:, :3])
+            frame = I1[0]
+            I1 = I1[0]
+
+        tmp_output = []
+        if ssim < 0.2:
+            for i in range((2**exp) - 1):
+                tmp_output.append(I0)
+
+        else:
+            tmp_output = make_inference(model, I0, I1, upscale_amount, 2**exp - 1) if exp else []
+
+        frame = pad_image(frame, upscale_amount)
+        tmp_output = [frame] + tmp_output
+        for i, frame in enumerate(tmp_output):
+            output.append(frame.to(output_device))
+    return output
+
+
+def load_rife_model(model_path):
+    model = Model()
+    model.load_model(model_path, -1)
+    model.eval()
+    return model
+
+
+# Create a generator that yields each frame, similar to cv2.VideoCapture
+def frame_generator(video_capture):
+    while True:
+        ret, frame = video_capture.read()
+        if not ret:
+            break
+        yield frame
+    video_capture.release()
+
+
+def rife_inference_with_path(model, video_path):
+    video_capture = cv2.VideoCapture(video_path)
+    tot_frame = video_capture.get(cv2.CAP_PROP_FRAME_COUNT)
+    pt_frame_data = []
+    pt_frame = skvideo.io.vreader(video_path)
+    for frame in pt_frame:
+        pt_frame_data.append(
+            torch.from_numpy(np.transpose(frame, (2, 0, 1))).to("cpu", non_blocking=True).float() / 255.0
+        )
+
+    pt_frame = torch.from_numpy(np.stack(pt_frame_data))
+    pt_frame = pt_frame.to(device)
+    pbar = utils.ProgressBar(tot_frame, desc="RIFE inference")
+    frames = ssim_interpolation_rife(model, pt_frame)
+    pt_image = torch.stack([frames[i].squeeze(0) for i in range(len(frames))])
+    image_np = VaeImageProcessor.pt_to_numpy(pt_image)  # (to [49, 512, 480, 3])
+    image_pil = VaeImageProcessor.numpy_to_pil(image_np)
+    video_path = utils.save_video(image_pil, fps=16)
+    if pbar:
+        pbar.update(1)
+    return video_path
+
+
+def rife_inference_with_latents(model, latents):
+    pbar = utils.ProgressBar(latents.shape[1], desc="RIFE inference")
+    rife_results = []
+    latents = latents.to(device)
+    for i in range(latents.size(0)):
+        #  [f, c, w, h]
+        latent = latents[i]
+
+        frames = ssim_interpolation_rife(model, latent)
+        pt_image = torch.stack([frames[i].squeeze(0) for i in range(len(frames))])  # (to [f, c, w, h])
+        rife_results.append(pt_image)
+
+    return torch.stack(rife_results)

inference/gradio_composite_demo/utils.py

@@ -0,0 +1,221 @@
+import math
+from typing import Union, List
+
+import torch
+import os
+from datetime import datetime
+import numpy as np
+import itertools
+import PIL.Image
+import safetensors.torch
+import tqdm
+import logging
+from diffusers.utils import export_to_video
+from spandrel import ModelLoader
+
+logger = logging.getLogger(__file__)
+
+
+def load_torch_file(ckpt, device=None, dtype=torch.float16):
+    if device is None:
+        device = torch.device("cpu")
+    if ckpt.lower().endswith(".safetensors") or ckpt.lower().endswith(".sft"):
+        sd = safetensors.torch.load_file(ckpt, device=device.type)
+    else:
+        if not "weights_only" in torch.load.__code__.co_varnames:
+            logger.warning(
+                "Warning torch.load doesn't support weights_only on this pytorch version, loading unsafely."
+            )
+
+        pl_sd = torch.load(ckpt, map_location=device, weights_only=True)
+        if "global_step" in pl_sd:
+            logger.debug(f"Global Step: {pl_sd['global_step']}")
+        if "state_dict" in pl_sd:
+            sd = pl_sd["state_dict"]
+        elif "params_ema" in pl_sd:
+            sd = pl_sd["params_ema"]
+        else:
+            sd = pl_sd
+
+    sd = {k: v.to(dtype) for k, v in sd.items()}
+    return sd
+
+
+def state_dict_prefix_replace(state_dict, replace_prefix, filter_keys=False):
+    if filter_keys:
+        out = {}
+    else:
+        out = state_dict
+    for rp in replace_prefix:
+        replace = list(
+            map(
+                lambda a: (a, "{}{}".format(replace_prefix[rp], a[len(rp) :])),
+                filter(lambda a: a.startswith(rp), state_dict.keys()),
+            )
+        )
+        for x in replace:
+            w = state_dict.pop(x[0])
+            out[x[1]] = w
+    return out
+
+
+def module_size(module):
+    module_mem = 0
+    sd = module.state_dict()
+    for k in sd:
+        t = sd[k]
+        module_mem += t.nelement() * t.element_size()
+    return module_mem
+
+
+def get_tiled_scale_steps(width, height, tile_x, tile_y, overlap):
+    return math.ceil((height / (tile_y - overlap))) * math.ceil((width / (tile_x - overlap)))
+
+
+@torch.inference_mode()
+def tiled_scale_multidim(
+    samples, function, tile=(64, 64), overlap=8, upscale_amount=4, out_channels=3, output_device="cpu", pbar=None
+):
+    dims = len(tile)
+    print(f"samples dtype:{samples.dtype}")
+    output = torch.empty(
+        [samples.shape[0], out_channels] + list(map(lambda a: round(a * upscale_amount), samples.shape[2:])),
+        device=output_device,
+    )
+
+    for b in range(samples.shape[0]):
+        s = samples[b : b + 1]
+        out = torch.zeros(
+            [s.shape[0], out_channels] + list(map(lambda a: round(a * upscale_amount), s.shape[2:])),
+            device=output_device,
+        )
+        out_div = torch.zeros(
+            [s.shape[0], out_channels] + list(map(lambda a: round(a * upscale_amount), s.shape[2:])),
+            device=output_device,
+        )
+
+        for it in itertools.product(*map(lambda a: range(0, a[0], a[1] - overlap), zip(s.shape[2:], tile))):
+            s_in = s
+            upscaled = []
+
+            for d in range(dims):
+                pos = max(0, min(s.shape[d + 2] - overlap, it[d]))
+                l = min(tile[d], s.shape[d + 2] - pos)
+                s_in = s_in.narrow(d + 2, pos, l)
+                upscaled.append(round(pos * upscale_amount))
+
+            ps = function(s_in).to(output_device)
+            mask = torch.ones_like(ps)
+            feather = round(overlap * upscale_amount)
+            for t in range(feather):
+                for d in range(2, dims + 2):
+                    m = mask.narrow(d, t, 1)
+                    m *= (1.0 / feather) * (t + 1)
+                    m = mask.narrow(d, mask.shape[d] - 1 - t, 1)
+                    m *= (1.0 / feather) * (t + 1)
+
+            o = out
+            o_d = out_div
+            for d in range(dims):
+                o = o.narrow(d + 2, upscaled[d], mask.shape[d + 2])
+                o_d = o_d.narrow(d + 2, upscaled[d], mask.shape[d + 2])
+
+            o += ps * mask
+            o_d += mask
+
+            if pbar is not None:
+                pbar.update(1)
+
+        output[b : b + 1] = out / out_div
+    return output
+
+
+def tiled_scale(
+    samples,
+    function,
+    tile_x=64,
+    tile_y=64,
+    overlap=8,
+    upscale_amount=4,
+    out_channels=3,
+    output_device="cpu",
+    pbar=None,
+):
+    return tiled_scale_multidim(
+        samples, function, (tile_y, tile_x), overlap, upscale_amount, out_channels, output_device, pbar
+    )
+
+
+def load_sd_upscale(ckpt, inf_device):
+    sd = load_torch_file(ckpt, device=inf_device)
+    if "module.layers.0.residual_group.blocks.0.norm1.weight" in sd:
+        sd = state_dict_prefix_replace(sd, {"module.": ""})
+    out = ModelLoader().load_from_state_dict(sd).half()
+    return out
+
+
+def upscale(upscale_model, tensor: torch.Tensor, inf_device, output_device="cpu") -> torch.Tensor:
+    memory_required = module_size(upscale_model.model)
+    memory_required += (
+        (512 * 512 * 3) * tensor.element_size() * max(upscale_model.scale, 1.0) * 384.0
+    )  # The 384.0 is an estimate of how much some of these models take, TODO: make it more accurate
+    memory_required += tensor.nelement() * tensor.element_size()
+    print(f"UPScaleMemory required: {memory_required / 1024 / 1024 / 1024} GB")
+
+    upscale_model.to(inf_device)
+    tile = 512
+    overlap = 32
+
+    steps = tensor.shape[0] * get_tiled_scale_steps(
+        tensor.shape[3], tensor.shape[2], tile_x=tile, tile_y=tile, overlap=overlap
+    )
+
+    pbar = ProgressBar(steps, desc="Tiling and Upscaling")
+
+    s = tiled_scale(
+        samples=tensor.to(torch.float16),
+        function=lambda a: upscale_model(a),
+        tile_x=tile,
+        tile_y=tile,
+        overlap=overlap,
+        upscale_amount=upscale_model.scale,
+        pbar=pbar,
+    )
+
+    upscale_model.to(output_device)
+    return s
+
+
+def upscale_batch_and_concatenate(upscale_model, latents, inf_device, output_device="cpu") -> torch.Tensor:
+    upscaled_latents = []
+    for i in range(latents.size(0)):
+        latent = latents[i]
+        upscaled_latent = upscale(upscale_model, latent, inf_device, output_device)
+        upscaled_latents.append(upscaled_latent)
+    return torch.stack(upscaled_latents)
+
+
+def save_video(tensor: Union[List[np.ndarray], List[PIL.Image.Image]], fps: int = 8):
+    timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
+    video_path = f"./output/{timestamp}.mp4"
+    os.makedirs(os.path.dirname(video_path), exist_ok=True)
+    export_to_video(tensor, video_path, fps=fps)
+    return video_path
+
+
+class ProgressBar:
+    def __init__(self, total, desc=None):
+        self.total = total
+        self.current = 0
+        self.b_unit = tqdm.tqdm(total=total, desc="ProgressBar context index: 0" if desc is None else desc)
+
+    def update(self, value):
+        if value > self.total:
+            value = self.total
+        self.current = value
+        if self.b_unit is not None:
+            self.b_unit.set_description("ProgressBar context index: {}".format(self.current))
+            self.b_unit.refresh()
+
+            # 更新进度
+            self.b_unit.update(self.current)

inference/gradio_web_demo.py

@@ -1,23 +1,21 @@
 import os
-import tempfile
 import threading
 import time
 
 import gradio as gr
-import numpy as np
 import torch
 from diffusers import CogVideoXPipeline
+from diffusers.utils import export_to_video
 from datetime import datetime, timedelta
 from openai import OpenAI
-import imageio
 import moviepy.editor as mp
-from typing import List, Union
-import PIL
 
-dtype = torch.bfloat16
+dtype = torch.float16
 device = "cuda" if torch.cuda.is_available() else "cpu"
-pipe = CogVideoXPipeline.from_pretrained("THUDM/CogVideoX-2b", torch_dtype=dtype)
+pipe = CogVideoXPipeline.from_pretrained("THUDM/CogVideoX-2b", torch_dtype=dtype).to(device)
-pipe.enable_model_cpu_offload()
+
+os.makedirs("./output", exist_ok=True)
+os.makedirs("./gradio_tmp", exist_ok=True)
 
 sys_prompt = """You are part of a team of bots that creates videos. You work with an assistant bot that will draw anything you say in square brackets.
 
@@ -33,25 +31,6 @@ Video descriptions must have the same num of words as examples below. Extra word
 """
 
 
-def export_to_video_imageio(
-    video_frames: Union[List[np.ndarray], List[PIL.Image.Image]], output_video_path: str = None, fps: int = 8
-) -> str:
-    """
-    Export the video frames to a video file using imageio lib to Avoid "green screen" issue (for example CogVideoX)
-    """
-    if output_video_path is None:
-        output_video_path = tempfile.NamedTemporaryFile(suffix=".mp4").name
-
-    if isinstance(video_frames[0], PIL.Image.Image):
-        video_frames = [np.array(frame) for frame in video_frames]
-
-    with imageio.get_writer(output_video_path, fps=fps) as writer:
-        for frame in video_frames:
-            writer.append_data(frame)
-
-    return output_video_path
-
-
 def convert_prompt(prompt: str, retry_times: int = 3) -> str:
     if not os.environ.get("OPENAI_API_KEY"):
         return prompt
@@ -104,22 +83,12 @@ def convert_prompt(prompt: str, retry_times: int = 3) -> str:
 
 def infer(prompt: str, num_inference_steps: int, guidance_scale: float, progress=gr.Progress(track_tqdm=True)):
     torch.cuda.empty_cache()
-
-    prompt_embeds, _ = pipe.encode_prompt(
-        prompt=prompt,
-        negative_prompt=None,
-        do_classifier_free_guidance=True,
-        num_videos_per_prompt=1,
-        max_sequence_length=226,
-        device=device,
-        dtype=dtype,
-    )
-
     video = pipe(
+        prompt=prompt,
+        num_videos_per_prompt=1,
         num_inference_steps=num_inference_steps,
+        num_frames=49,
         guidance_scale=guidance_scale,
-        prompt_embeds=prompt_embeds,
-        negative_prompt_embeds=torch.zeros_like(prompt_embeds),
     ).frames[0]
 
     return video
@@ -129,7 +98,7 @@ def save_video(tensor):
     timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
     video_path = f"./output/{timestamp}.mp4"
     os.makedirs(os.path.dirname(video_path), exist_ok=True)
-    export_to_video_imageio(tensor[1:], video_path)
+    export_to_video(tensor, video_path)
     return video_path
 
 
@@ -146,14 +115,16 @@ def delete_old_files():
     while True:
         now = datetime.now()
         cutoff = now - timedelta(minutes=10)
-        output_dir = "./output"
+        directories = ["./output", "./gradio_tmp"]
-        for filename in os.listdir(output_dir):
+
-            file_path = os.path.join(output_dir, filename)
+        for directory in directories:
+            for filename in os.listdir(directory):
+                file_path = os.path.join(directory, filename)
                 if os.path.isfile(file_path):
                     file_mtime = datetime.fromtimestamp(os.path.getmtime(file_path))
                     if file_mtime < cutoff:
                         os.remove(file_path)
-        time.sleep(600)  # Sleep for 10 minutes
+        time.sleep(600)
 
 
 threading.Thread(target=delete_old_files, daemon=True).start()
@@ -164,8 +135,9 @@ with gr.Blocks() as demo:
                CogVideoX-2B Huggingface Space🤗
            </div>
            <div style="text-align: center;">
-               <a href="https://huggingface.co/THUDM/CogVideoX-2b">🤗 Model Hub</a> |
+               <a href="https://huggingface.co/THUDM/CogVideoX-2B">🤗 2B Model Hub</a> |
-               <a href="https://github.com/THUDM/CogVideo">🌐 Github</a> 
+               <a href="https://github.com/THUDM/CogVideo">🌐 Github</a> |
+               <a href="https://arxiv.org/pdf/2408.06072">📜 arxiv </a>
            </div>
 
            <div style="text-align: center; font-size: 15px; font-weight: bold; color: red; margin-bottom: 20px;">
@@ -176,18 +148,17 @@ with gr.Blocks() as demo:
     with gr.Row():
         with gr.Column():
             prompt = gr.Textbox(label="Prompt (Less than 200 Words)", placeholder="Enter your prompt here", lines=5)
+
             with gr.Row():
                 gr.Markdown(
-                    "✨Upon pressing the enhanced prompt button, we will use [GLM-4 Model](https://github.com/THUDM/GLM-4) to polish the prompt and overwrite the original one."
+                    "✨Upon pressing the enhanced prompt button, we will use [GLM-4 Model](https://github.com/THUDM/GLM-4) to polish the prompt and overwrite the original one.")
-                )
                 enhance_button = gr.Button("✨ Enhance Prompt(Optional)")
 
             with gr.Column():
-                gr.Markdown(
+                gr.Markdown("**Optional Parameters** (default values are recommended)<br>"
-                    "**Optional Parameters** (default values are recommended)<br>"
+                            "Increasing the number of inference steps will produce more detailed videos, but it will slow down the process.<br>"
-                    "Turn Inference Steps larger if you want more detailed video, but it will be slower.<br>"
+                            "50 steps are recommended for most cases.<br>"
-                    "50 steps are recommended for most cases. will cause 120 seconds for inference.<br>"
+                            "For the 5B model, 50 steps will take approximately 350 seconds.")
-                )
                 with gr.Row():
                     num_inference_steps = gr.Number(label="Inference Steps", value=50)
                     guidance_scale = gr.Number(label="Guidance Scale", value=6.0)
@@ -200,41 +171,43 @@ with gr.Blocks() as demo:
                 download_gif_button = gr.File(label="📥 Download GIF", visible=False)
 
     gr.Markdown("""
-        <table border="1" style="width: 100%; text-align: left; margin-top: 20px;">
+    <table border="0" style="width: 100%; text-align: left; margin-top: 20px;">
+         <div style="text-align: center; font-size: 24px; font-weight: bold; margin-bottom: 20px;">
+               Demo Videos with 50 Inference Steps and 6.0 Guidance Scale.
+         </div>
         <tr>
-                <th>Prompt</th>
+            <td style="width: 25%; vertical-align: top; font-size: 0.8em;">
-                <th>Video URL</th>
+                <p>A detailed wooden toy ship with intricately carved masts and sails is seen gliding smoothly over a plush, blue carpet that mimics the waves of the sea. The ship's hull is painted a rich brown, with tiny windows. The carpet, soft and textured, provides a perfect backdrop, resembling an oceanic expanse. Surrounding the ship are various other toys and children's items, hinting at a playful environment. The scene captures the innocence and imagination of childhood, with the toy ship's journey symbolizing endless adventures in a whimsical, indoor setting.</p>
-                <th>Inference Steps</th>
+            </td>
-                <th>Guidance Scale</th>
+            <td style="width: 25%; vertical-align: top;">
+                <video src="https://github.com/user-attachments/assets/ea3af39a-3160-4999-90ec-2f7863c5b0e9" width="100%" controls autoplay></video>
+            </td>
+            <td style="width: 25%; vertical-align: top; font-size: 0.8em;">
+                <p>The camera follows behind a white vintage SUV with a black roof rack as it speeds up a steep dirt road surrounded by pine trees on a steep mountain slope, dust kicks up from its tires, the sunlight shines on the SUV as it speeds along the dirt road, casting a warm glow over the scene. The dirt road curves gently into the distance, with no other cars or vehicles in sight. The trees on either side of the road are redwoods, with patches of greenery scattered throughout. The car is seen from the rear following the curve with ease, making it seem as if it is on a rugged drive through the rugged terrain. The dirt road itself is surrounded by steep hills and mountains, with a clear blue sky above with wispy clouds.</p>
+            </td>
+            <td style="width: 25%; vertical-align: top;">
+                <video src="https://github.com/user-attachments/assets/9de41efd-d4d1-4095-aeda-246dd834e91d" width="100%" controls autoplay></video>
+            </td>
         </tr>
         <tr>
-                <td>A detailed wooden toy ship with intricately carved masts and sails is seen gliding smoothly over a plush, blue carpet that mimics the waves of the sea. The ship's hull is painted a rich brown, with tiny windows. The carpet, soft and textured, provides a perfect backdrop, resembling an oceanic expanse. Surrounding the ship are various other toys and children's items, hinting at a playful environment. The scene captures the innocence and imagination of childhood, with the toy ship's journey symbolizing endless adventures in a whimsical, indoor setting.</td>
+            <td style="width: 25%; vertical-align: top; font-size: 0.8em;">
-                <td><a href="https://github.com/THUDM/CogVideo/raw/main/resources/videos/1.mp4">Video 1</a></td>
+                <p>A street artist, clad in a worn-out denim jacket and a colorful bandana, stands before a vast concrete wall in the heart, holding a can of spray paint, spray-painting a colorful bird on a mottled wall.</p>
-                <td>50</td>
+            </td>
-                <td>6</td>
+            <td style="width: 25%; vertical-align: top;">
-            </tr>
+                <video src="https://github.com/user-attachments/assets/941d6661-6a8d-4a1b-b912-59606f0b2841" width="100%" controls autoplay></video>
-            <tr>
+            </td>
-                <td>The camera follows behind a white vintage SUV with a black roof rack as it speeds up a steep dirt road surrounded by pine trees on a steep mountain slope, dust kicks up from it’s tires, the sunlight shines on the SUV as it speeds along the dirt road, casting a warm glow over the scene. The dirt road curves gently into the distance, with no other cars or vehicles in sight. The trees on either side of the road are redwoods, with patches of greenery scattered throughout. The car is seen from the rear following the curve with ease, making it seem as if it is on a rugged drive through the rugged terrain. The dirt road itself is surrounded by steep hills and mountains, with a clear blue sky above with wispy clouds.</td>
+            <td style="width: 25%; vertical-align: top; font-size: 0.8em;">
-                <td><a href="https://github.com/THUDM/CogVideo/raw/main/resources/videos/2.mp4">Video 2</a></td>
+                <p>In the haunting backdrop of a war-torn city, where ruins and crumbled walls tell a story of devastation, a poignant close-up frames a young girl. Her face is smudged with ash, a silent testament to the chaos around her. Her eyes glistening with a mix of sorrow and resilience, capturing the raw emotion of a world that has lost its innocence to the ravages of conflict.</p>
-                <td>50</td>
+            </td>
-                <td>6</td>
+            <td style="width: 25%; vertical-align: top;">
-            </tr>
+                <video src="https://github.com/user-attachments/assets/938529c4-91ae-4f60-b96b-3c3947fa63cb" width="100%" controls autoplay></video>
-            <tr>
+            </td>
-                <td>A street artist, clad in a worn-out denim jacket and a colorful bandana, stands before a vast concrete wall in the heart, holding a can of spray paint, spray-painting a colorful bird on a mottled wall.</td>
-                <td><a href="https://github.com/THUDM/CogVideo/raw/main/resources/videos/3.mp4">Video 3</a></td>
-                <td>50</td>
-                <td>6</td>
-            </tr>
-            <tr>
-                <td>In the haunting backdrop of a war-torn city, where ruins and crumbled walls tell a story of devastation, a poignant close-up frames a young girl. Her face is smudged with ash, a silent testament to the chaos around her. Her eyes glistening with a mix of sorrow and resilience, capturing the raw emotion of a world that has lost its innocence to the ravages of conflict.</td>
-                <td><a href="https://github.com/THUDM/CogVideo/raw/main/resources/videos/4.mp4">Video 4</a></td>
-                <td>50</td>
-                <td>6</td>
         </tr>
     </table>
     """)
 
-    def generate(prompt, num_inference_steps, guidance_scale, progress=gr.Progress(track_tqdm=True)):
+
+    def generate(prompt, num_inference_steps, guidance_scale, model_choice, progress=gr.Progress(track_tqdm=True)):
         tensor = infer(prompt, num_inference_steps, guidance_scale, progress=progress)
         video_path = save_video(tensor)
         video_update = gr.update(visible=True, value=video_path)
@@ -243,16 +216,22 @@ with gr.Blocks() as demo:
 
         return video_path, video_update, gif_update
 
+
     def enhance_prompt_func(prompt):
         return convert_prompt(prompt, retry_times=1)
 
+
     generate_button.click(
         generate,
         inputs=[prompt, num_inference_steps, guidance_scale],
-        outputs=[video_output, download_video_button, download_gif_button],
+        outputs=[video_output, download_video_button, download_gif_button]
     )
 
-    enhance_button.click(enhance_prompt_func, inputs=[prompt], outputs=[prompt])
+    enhance_button.click(
+        enhance_prompt_func,
+        inputs=[prompt],
+        outputs=[prompt]
+    )
 
 if __name__ == "__main__":
-    demo.launch(server_name="127.0.0.1", server_port=7870, share=True)
+    demo.launch()

requirements.txt

@@ -1,14 +1,15 @@
-git+https://github.com/huggingface/diffusers.git@main#egg=diffusers
+diffusers>=0.30.1 #git+https://github.com/huggingface/diffusers.git@main#egg=diffusers is suggested
-transformers==4.44.0
+transformers>=4.44.0  # The development team is working on version 0.44.2
-git+https://github.com/huggingface/accelerate.git@main#egg=accelerate
+accelerate>=0.33.0 #git+https://github.com/huggingface/accelerate.git@main#egg=accelerate is suggested
-sentencepiece==0.2.0 # T5
+sentencepiece>=0.2.0 # T5 used
-SwissArmyTransformer==0.4.12 # Inference
+SwissArmyTransformer>=0.4.12
-torch==2.4.0 # Tested in 2.2 2.3 2.4 and 2.5
+numpy==1.26.0
-torchvision==0.19.0
+torch>=2.4.0 # Tested in 2.2 2.3 2.4 and 2.5, The development team is working on version 2.4.0.
-gradio==4.40.0 # For HF gradio demo
+torchvision>=0.19.0 # The development team is working on version 0.19.0.
-pillow==9.5.0 # For HF gradio demo
+gradio>=4.42.0 # For HF gradio demo
-streamlit==1.37.0 # For streamlit web demo
+streamlit>=1.37.1 # For streamlit web demo
 imageio==2.34.2 # For diffusers inference export video
 imageio-ffmpeg==0.5.1 # For diffusers inference export video
-openai==1.40.6 # For prompt refiner
+openai>=1.42.0 # For prompt refiner
 moviepy==1.0.3 # For export video
+pillow==9.5.0

resources/galary_prompt.md

@@ -0,0 +1,31 @@
+## CogVideoX-5B
+
+Videos 1-8:
+
+1. A garden comes to life as a kaleidoscope of butterflies flutters amidst the blossoms, their delicate wings casting shadows on the petals below. In the background, a grand fountain cascades water with a gentle splendor, its rhythmic sound providing a soothing backdrop. Beneath the cool shade of a mature tree, a solitary wooden chair invites solitude and reflection, its smooth surface worn by the touch of countless visitors seeking a moment of tranquility in nature's embrace.
+
+2. A small boy, head bowed and determination etched on his face, sprints through the torrential downpour as闪电 crackles and 雷鸣 rumbles in the distance. The relentless rain pounds the ground, creating a chaotic dance of water droplets that mirror the Dramatic sky's anger. In the far background, the silhouette of a cozy home beckons, a faint beacon of safety and warmth amidst the fierce weather. The scene is one of perseverance and the unyielding spirit of a child braving the elements.
+
+3. A suited astronaut, with the red dust of Mars clinging to their boots, reaches out to shake hands with an alien being, their skin a shimmering blue, under the pink-tinged sky of the fourth planet. In the background, a sleek silver rocket, a beacon of human ingenuity, stands tall, its engines powered down, as the two representatives of different worlds exchange a historic greeting amidst the desolate beauty of the Martian landscape.
+
+4. An elderly gentleman, with a serene expression, sits at the water's edge, a steaming cup of tea by his side. He is engrossed in his artwork, brush in hand, as he renders an oil painting on a canvas that's propped up against a small, weathered table. The sea breeze whispers through his silver hair, gently billowing his loose-fitting white shirt, while the salty air adds an intangible element to his masterpiece in progress. The scene is one of tranquility and inspiration, with the artist's canvas capturing the vibrant hues of the setting sun reflecting off the tranquil sea.
+
+5. In a dimly lit bar, purplish light bathes the face of a mature man, his eyes blinking thoughtfully as he ponders in close-up, the background artfully blurred to focus on his introspective expression, the ambiance of the bar a mere suggestion of shadows and soft lighting.
+
+6. A golden retriever, sporting sleek black sunglasses, with its lengthy fur flowing in the breeze, sprints playfully across a rooftop terrace, recently refreshed by a light rain. The scene unfolds from a distance, the dog's energetic bounds growing larger as it approaches the camera, its tail wagging with unrestrained joy, while droplets of water glisten on the concrete behind it. The overcast sky provides a dramatic backdrop, emphasizing the vibrant golden coat of the canine as it dashes towards the viewer.
+
+7. On a brilliant sunny day, the lakeshore is lined with an array of willow trees, their slender branches swaying gently in the soft breeze. The tranquil surface of the lake reflects the clear blue sky, while several elegant swans glide gracefully through the still water, leaving behind delicate ripples that disturb the mirror-like quality of the lake. The scene is one of serene beauty, with the willows' greenery providing a picturesque frame for the peaceful avian visitors.
+
+8. A Chinese mother, draped in a soft, pastel-colored robe, gently rocks back and forth in a cozy rocking chair positioned in the tranquil setting of a nursery. The dimly lit bedroom is adorned with whimsical mobiles dangling from the ceiling, casting shadows that dance on the walls. Her baby, swaddled in a delicate, patterned blanket, rests against her chest, the child's earlier cries now replaced by contented coos as the mother's soothing voice lulls the little one to sleep. The scent of lavender fills the air, adding to the serene atmosphere, while a warm, orange glow from a nearby nightlight illuminates the scene with a gentle hue, capturing a moment of tender love and comfort.
+
+## CogVideoX-2B
+
+Videos 1-4: 
+
+1. A detailed wooden toy ship with intricately carved masts and sails is seen gliding smoothly over a plush, blue carpet that mimics the waves of the sea. The ship's hull is painted a rich brown, with tiny windows. The carpet, soft and textured, provides a perfect backdrop, resembling an oceanic expanse. Surrounding the ship are various other toys and children's items, hinting at a playful environment. The scene captures the innocence and imagination of childhood, with the toy ship's journey symbolizing endless adventures in a whimsical, indoor setting.
+
+2. The camera follows behind a white vintage SUV with a black roof rack as it speeds up a steep dirt road surrounded by pine trees on a steep mountain slope, dust kicks up from its tires, the sunlight shines on the SUV as it speeds along the dirt road, casting a warm glow over the scene. The dirt road curves gently into the distance, with no other cars or vehicles in sight. The trees on either side of the road are redwoods, with patches of greenery scattered throughout. The car is seen from the rear following the curve with ease, making it seem as if it is on a rugged drive through the rugged terrain. The dirt road itself is surrounded by steep hills and mountains, with a clear blue sky above with wispy clouds.
+
+3. A street artist, clad in a worn-out denim jacket and a colorful bandana, stands before a vast concrete wall in the heart, holding a can of spray paint, spray-painting a colorful bird on a mottled wall.
+
+4. In the haunting backdrop of a war-torn city, where ruins and crumbled walls tell a story of devastation, a poignant close-up frames a young girl. Her face is smudged with ash, a silent testament to the chaos around her. Her eyes glistening with a mix of sorrow and resilience, capturing the raw emotion of a world that has lost its innocence to the ravages of conflict.

sat/configs/cogvideox_5b.yaml

@@ -0,0 +1,153 @@
+model:
+  scale_factor: 0.7 # different from cogvideox_2b_infer.yaml
+  disable_first_stage_autocast: true
+  log_keys:
+    - txt
+  
+  denoiser_config:
+    target: sgm.modules.diffusionmodules.denoiser.DiscreteDenoiser
+    params:
+      num_idx: 1000
+      quantize_c_noise: False
+
+      weighting_config:
+        target: sgm.modules.diffusionmodules.denoiser_weighting.EpsWeighting
+      scaling_config:
+        target: sgm.modules.diffusionmodules.denoiser_scaling.VideoScaling
+      discretization_config:
+        target: sgm.modules.diffusionmodules.discretizer.ZeroSNRDDPMDiscretization
+        params:
+          shift_scale: 1.0 # different from cogvideox_2b_infer.yaml
+
+  network_config:
+    target: dit_video_concat.DiffusionTransformer
+    params:
+      time_embed_dim: 512
+      elementwise_affine: True
+      num_frames: 49
+      time_compressed_rate: 4
+      latent_width: 90
+      latent_height: 60
+      num_layers: 42 # different from cogvideox_2b_infer.yaml
+      patch_size: 2
+      in_channels: 16
+      out_channels: 16
+      hidden_size: 3072 # different from cogvideox_2b_infer.yaml
+      adm_in_channels: 256
+      num_attention_heads: 48 # different from cogvideox_2b_infer.yaml
+
+      transformer_args:
+        checkpoint_activations: True
+        vocab_size: 1
+        max_sequence_length: 64
+        layernorm_order: pre
+        skip_init: false
+        model_parallel_size: 1
+        is_decoder: false
+
+      modules:
+        pos_embed_config:
+          target: dit_video_concat.Rotary3DPositionEmbeddingMixin # different from cogvideox_2b_infer.yaml
+          params:
+            hidden_size_head: 64
+            text_length: 226
+
+        patch_embed_config:
+          target: dit_video_concat.ImagePatchEmbeddingMixin
+          params:
+            text_hidden_size: 4096
+
+        adaln_layer_config:
+          target: dit_video_concat.AdaLNMixin
+          params:
+            qk_ln: True
+
+        final_layer_config:
+          target: dit_video_concat.FinalLayerMixin
+
+  conditioner_config:
+    target: sgm.modules.GeneralConditioner
+    params:
+      emb_models:
+        - is_trainable: false
+          input_key: txt
+          ucg_rate: 0.1
+          target: sgm.modules.encoders.modules.FrozenT5Embedder
+          params:
+            model_dir: "t5-v1_1-xxl"
+            max_length: 226
+
+  first_stage_config:
+    target: vae_modules.autoencoder.VideoAutoencoderInferenceWrapper
+    params:
+      cp_size: 1
+      ckpt_path: "cogvideox-5b-sat/vae/3d-vae.pt"
+      ignore_keys: [ 'loss' ]
+
+      loss_config:
+        target: torch.nn.Identity
+
+      regularizer_config:
+        target: vae_modules.regularizers.DiagonalGaussianRegularizer
+
+      encoder_config:
+        target: vae_modules.cp_enc_dec.ContextParallelEncoder3D
+        params:
+          double_z: true
+          z_channels: 16
+          resolution: 256
+          in_channels: 3
+          out_ch: 3
+          ch: 128
+          ch_mult: [ 1, 2, 2, 4 ]
+          attn_resolutions: [ ]
+          num_res_blocks: 3
+          dropout: 0.0
+          gather_norm: True
+
+      decoder_config:
+        target: vae_modules.cp_enc_dec.ContextParallelDecoder3D
+        params:
+          double_z: True
+          z_channels: 16
+          resolution: 256
+          in_channels: 3
+          out_ch: 3
+          ch: 128
+          ch_mult: [ 1, 2, 2, 4 ]
+          attn_resolutions: [ ]
+          num_res_blocks: 3
+          dropout: 0.0
+          gather_norm: False
+
+  loss_fn_config:
+    target: sgm.modules.diffusionmodules.loss.VideoDiffusionLoss
+    params:
+      offset_noise_level: 0
+      sigma_sampler_config:
+        target: sgm.modules.diffusionmodules.sigma_sampling.DiscreteSampling
+        params:
+          uniform_sampling: True
+          num_idx: 1000
+          discretization_config:
+            target: sgm.modules.diffusionmodules.discretizer.ZeroSNRDDPMDiscretization
+            params:
+              shift_scale: 1.0 # different from cogvideox_2b_infer.yaml
+
+  sampler_config:
+    target: sgm.modules.diffusionmodules.sampling.VPSDEDPMPP2MSampler
+    params:
+      num_steps: 50
+      verbose: True
+
+      discretization_config:
+        target: sgm.modules.diffusionmodules.discretizer.ZeroSNRDDPMDiscretization
+        params:
+          shift_scale: 1.0 # different from cogvideox_2b_infer.yaml
+
+      guider_config:
+        target: sgm.modules.diffusionmodules.guiders.DynamicCFG
+        params:
+          scale: 6
+          exp: 5
+          num_steps: 50

sat/configs/cogvideox_5b_lora.yaml

@@ -0,0 +1,159 @@
+model:
+  scale_factor: 0.7 # different from cogvideox_2b_infer.yaml
+  disable_first_stage_autocast: true
+  not_trainable_prefixes: ['all'] ## Using Lora
+  log_keys:
+    - txt
+
+  denoiser_config:
+    target: sgm.modules.diffusionmodules.denoiser.DiscreteDenoiser
+    params:
+      num_idx: 1000
+      quantize_c_noise: False
+
+      weighting_config:
+        target: sgm.modules.diffusionmodules.denoiser_weighting.EpsWeighting
+      scaling_config:
+        target: sgm.modules.diffusionmodules.denoiser_scaling.VideoScaling
+      discretization_config:
+        target: sgm.modules.diffusionmodules.discretizer.ZeroSNRDDPMDiscretization
+        params:
+          shift_scale: 1.0 # different from cogvideox_2b_infer.yaml
+
+  network_config:
+    target: dit_video_concat.DiffusionTransformer
+    params:
+      time_embed_dim: 512
+      elementwise_affine: True
+      num_frames: 49
+      time_compressed_rate: 4
+      latent_width: 90
+      latent_height: 60
+      num_layers: 42 # different from cogvideox_2b_infer.yaml
+      patch_size: 2
+      in_channels: 16
+      out_channels: 16
+      hidden_size: 3072 # different from cogvideox_2b_infer.yaml
+      adm_in_channels: 256
+      num_attention_heads: 48 # different from cogvideox_2b_infer.yaml
+
+      transformer_args:
+        checkpoint_activations: True
+        vocab_size: 1
+        max_sequence_length: 64
+        layernorm_order: pre
+        skip_init: false
+        model_parallel_size: 1
+        is_decoder: false
+
+      modules:
+        pos_embed_config:
+          target: dit_video_concat.Rotary3DPositionEmbeddingMixin # different from cogvideox_2b_infer.yaml
+          params:
+            hidden_size_head: 64
+            text_length: 226
+
+        lora_config: ## Using Lora
+          target: sat.model.finetune.lora2.LoraMixin
+          params:
+            r: 128
+
+        patch_embed_config:
+          target: dit_video_concat.ImagePatchEmbeddingMixin
+          params:
+            text_hidden_size: 4096
+
+        adaln_layer_config:
+          target: dit_video_concat.AdaLNMixin
+          params:
+            qk_ln: True
+
+        final_layer_config:
+          target: dit_video_concat.FinalLayerMixin
+
+  conditioner_config:
+    target: sgm.modules.GeneralConditioner
+    params:
+      emb_models:
+        - is_trainable: false
+          input_key: txt
+          ucg_rate: 0.1
+          target: sgm.modules.encoders.modules.FrozenT5Embedder
+          params:
+            model_dir: "t5-v1_1-xxl"
+            max_length: 226
+
+  first_stage_config:
+    target: vae_modules.autoencoder.VideoAutoencoderInferenceWrapper
+    params:
+      cp_size: 1
+      ckpt_path: "cogvideox-5b-sat/vae/3d-vae.pt"
+      ignore_keys: [ 'loss' ]
+
+      loss_config:
+        target: torch.nn.Identity
+
+      regularizer_config:
+        target: vae_modules.regularizers.DiagonalGaussianRegularizer
+
+      encoder_config:
+        target: vae_modules.cp_enc_dec.ContextParallelEncoder3D
+        params:
+          double_z: true
+          z_channels: 16
+          resolution: 256
+          in_channels: 3
+          out_ch: 3
+          ch: 128
+          ch_mult: [ 1, 2, 2, 4 ]
+          attn_resolutions: [ ]
+          num_res_blocks: 3
+          dropout: 0.0
+          gather_norm: True
+
+      decoder_config:
+        target: vae_modules.cp_enc_dec.ContextParallelDecoder3D
+        params:
+          double_z: True
+          z_channels: 16
+          resolution: 256
+          in_channels: 3
+          out_ch: 3
+          ch: 128
+          ch_mult: [ 1, 2, 2, 4 ]
+          attn_resolutions: [ ]
+          num_res_blocks: 3
+          dropout: 0.0
+          gather_norm: False
+
+  loss_fn_config:
+    target: sgm.modules.diffusionmodules.loss.VideoDiffusionLoss
+    params:
+      offset_noise_level: 0
+      sigma_sampler_config:
+        target: sgm.modules.diffusionmodules.sigma_sampling.DiscreteSampling
+        params:
+          uniform_sampling: True
+          num_idx: 1000
+          discretization_config:
+            target: sgm.modules.diffusionmodules.discretizer.ZeroSNRDDPMDiscretization
+            params:
+              shift_scale: 1.0 # different from cogvideox_2b_infer.yaml
+
+  sampler_config:
+    target: sgm.modules.diffusionmodules.sampling.VPSDEDPMPP2MSampler
+    params:
+      num_steps: 50
+      verbose: True
+
+      discretization_config:
+        target: sgm.modules.diffusionmodules.discretizer.ZeroSNRDDPMDiscretization
+        params:
+          shift_scale: 1.0
+
+      guider_config:
+        target: sgm.modules.diffusionmodules.guiders.DynamicCFG
+        params:
+          scale: 6
+          exp: 5
+          num_steps: 50

sat/configs/inference.yaml

@@ -1,7 +1,7 @@
 args:
   latent_channels: 16
   mode: inference
-  # load: "{your_CogVideoX-2b-sat_path}/transformer" # This is for Full model without lora adapter
+  load: "{your CogVideoX SAT folder}/transformer" # This is for Full model without lora adapter
   # load: "{your lora folder} such as zRzRzRzRzRzRzR/lora-disney-08-20-13-28" # This is for Full model without lora adapter
 
   batch_size: 1

sat/data_video.py

@@ -147,7 +147,9 @@ def pad_last_frame(tensor, num_frames):
     # T, H, W, C
     if len(tensor) < num_frames:
         pad_length = num_frames - len(tensor)
-        pad_tensor = torch.zeros([pad_length, *tensor.shape[1:]], dtype=tensor.dtype, device=tensor.device)
+        # Use the last frame to pad instead of zero
+        last_frame = tensor[-1]
+        pad_tensor = last_frame.unsqueeze(0).expand(pad_length, *tensor.shape[1:])
         padded_tensor = torch.cat([tensor, pad_tensor], dim=0)
         return padded_tensor
     else:

sat/inference.sh

@@ -4,7 +4,7 @@ echo "CUDA_VISIBLE_DEVICES=$CUDA_VISIBLE_DEVICES"
 
 environs="WORLD_SIZE=1 RANK=0 LOCAL_RANK=0 LOCAL_WORLD_SIZE=1"
 
-run_cmd="$environs python sample_video.py --base configs/cogvideox_2b.yaml configs/inference.yaml --seed $RANDOM"
+run_cmd="$environs python sample_video.py --base configs/cogvideox_5b.yaml configs/inference.yaml --seed $RANDOM"
 
 echo ${run_cmd}
 eval ${run_cmd}