JLNTV_IPD/CogVideo
4
0
Fork 0
mirror of https://github.com/THUDM/CogVideo.git synced 2025-04-05 19:41:59 +08:00
Code Issues Packages Projects Releases Wiki Activity

Merge pull request #272 from THUDM/CogVideoX_dev

Browse Source 
discord link update
This commit is contained in:
Yuxuan.Zhang 2024-09-12 10:12:20 +08:00 committed by GitHub
commit 87ad61b531
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
10 changed files with 213 additions and 87 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
README.md
View File

@ -14,7 +14,7 @@ Experience the CogVideoX-5B model online at <a href="https://huggingface.co/spac
📚 View the <a href="https://arxiv.org/abs/2408.06072" target="_blank">paper</a> and <a href="https://zhipu-ai.feishu.cn/wiki/DHCjw1TrJiTyeukfc9RceoSRnCh" target="_blank">user guide</a>
</p>
<p align="center">
👋 Join our <a href="resources/WECHAT.md" target="_blank">WeChat</a> and <a href="https://discord.gg/B94UfuhN" target="_blank">Discord</a>
👋 Join our <a href="resources/WECHAT.md" target="_blank">WeChat</a> and <a href="https://discord.gg/Qqjtj69f" target="_blank">Discord</a>
</p>
<p align="center">
📍 Visit <a href="https://chatglm.cn/video?lang=en?fr=osm_cogvideo">QingYing</a> and <a href="https://open.bigmodel.cn/?utm_campaign=open&_channel_track_key=OWTVNma9">API Platform</a> to experience larger-scale commercial video generation models.
@ -279,10 +279,11 @@ 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.
+ [ComfyUI-CogVideoXWrapper](https://github.com/kijai/ComfyUI-CogVideoXWrapper) Use the ComfyUI framework to integrate CogVideoX into your workflow.
+ [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.
+ [AutoDL Image](https://www.codewithgpu.com/i/THUDM/CogVideo/CogVideoX-5b-demo): A one-click deployment Huggingface
+ [AutoDL Space](https://www.codewithgpu.com/i/THUDM/CogVideo/CogVideoX-5b-demo): A one-click deployment Huggingface
Space image provided by community members.
+ [Colab Space](https://github.com/camenduru/CogVideoX-5B-jupyter) Run the CogVideoX-5B model using Jupyter Notebook on
Colab.
@ -336,9 +337,6 @@ This folder contains some tools for model conversion / caption generation, etc.
+ [convert_weight_sat2hf](tools/convert_weight_sat2hf.py): Convert SAT model weights to Huggingface model weights.
+ [caption_demo](tools/caption): Caption tool, a model that understands videos and outputs them in text.
+ [AutoDL Mirror](https://www.codewithgpu.com/i/THUDM/CogVideo/CogVideoX-5b-demo): A one-click deployment of Huggingface
Space mirror provided by community members.
## CogVideo(ICLR'23)
The official repo for the

3
README_ja.md
View File

@ -14,7 +14,7 @@
📚 <a href="https://arxiv.org/abs/2408.06072" target="_blank">論文</a>と<a href="https://zhipu-ai.feishu.cn/wiki/DHCjw1TrJiTyeukfc9RceoSRnCh" target="_blank">使用ドキュメント</a>を表示します。
</p>
<p align="center">
👋 <a href="resources/WECHAT.md" target="_blank">WeChat</a> と <a href="https://discord.gg/B94UfuhN" target="_blank">Discord</a> に参加
👋 <a href="resources/WECHAT.md" target="_blank">WeChat</a> と <a href="https://discord.gg/Qqjtj69f" target="_blank">Discord</a> に参加
</p>
<p align="center">
📍 <a href="https://chatglm.cn/video?lang=en?fr=osm_cogvideo">清影</a> と <a href="https://open.bigmodel.cn/?utm_campaign=open&_channel_track_key=OWTVNma9">APIプラットフォーム</a> を訪問して、より大規模な商用ビデオ生成モデルを体験
@ -261,6 +261,7 @@ pipe.vae.enable_tiling()
+ [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。
+ + [ComfyUI-CogVideoXWrapper](https://github.com/kijai/ComfyUI-CogVideoXWrapper) ComfyUIフレームワークを使用して、CogVideoXをワークフローに統合します。
+ [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。
+ [AutoDLイメージ](https://www.codewithgpu.com/i/THUDM/CogVideo/CogVideoX-5b-demo): コミュニティメンバーが提供するHuggingface
Spaceイメージのワンクリックデプロイメント。

3
README_zh.md
View File

@ -15,7 +15,7 @@
📚 查看 <a href="https://arxiv.org/abs/2408.06072" target="_blank">论文</a> 和 <a href="https://zhipu-ai.feishu.cn/wiki/DHCjw1TrJiTyeukfc9RceoSRnCh" target="_blank">使用文档</a>
</p>
<p align="center">
👋 加入我们的 <a href="resources/WECHAT.md" target="_blank">微信</a> 和 <a href="https://discord.gg/B94UfuhN" target="_blank">Discord</a>
👋 加入我们的 <a href="resources/WECHAT.md" target="_blank">微信</a> 和 <a href="https://discord.gg/Qqjtj69f" target="_blank">Discord</a>
</p>
<p align="center">
📍 前往<a href="https://chatglm.cn/video?fr=osm_cogvideox"> 清影</a> 和 <a href="https://open.bigmodel.cn/?utm_campaign=open&_channel_track_key=OWTVNma9"> API平台</a> 体验更大规模的商业版视频生成模型。
@ -249,6 +249,7 @@ pipe.vae.enable_tiling()
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。
+ [ComfyUI-CogVideoXWrapper](https://github.com/kijai/ComfyUI-CogVideoXWrapper) 使用ComfyUI框架，将CogVideoX加入到你的工作流中。
+ [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。
+ [AutoDL镜像](https://www.codewithgpu.com/i/THUDM/CogVideo/CogVideoX-5b-demo): 由社区成员提供的一键部署Huggingface
Space镜像。

1
inference/cli_demo.py
View File

@ -53,7 +53,6 @@ def generate_video(
# 3. Enable CPU offload for the model, enable tiling.
# turn off if you have multiple GPUs or enough GPU memory(such as H100) and it will cost less time in inference
pipe.enable_model_cpu_offload()
pipe.enable_sequential_cpu_offload()
pipe.vae.enable_slicing()
pipe.vae.enable_tiling()

93
inference/convert_demo.py
View File

@ -1,22 +1,27 @@
"""
The CogVideoX model is designed to generate high-quality videos based on detailed and highly descriptive prompts.
The model performs best when provided with refined, granular prompts, which enhance the quality of video generation.
This script is designed to assist with transforming simple user inputs into detailed prompts suitable for CogVideoX.
It can handle both text-to-video (t2v) and image-to-video (i2v) conversions.
The CogVideoX model is pre-trained and fine-tuned using longer and more detailed prompts.Therefore, it requires highly granular and detailed prompts as input.This script aims to transform user inputs into executable inputs for CogVideoX, enabling superior video generation.
- For text-to-video, simply provide the prompt.
- For image-to-video, provide the path to the image file and an optional user input.
The image will be encoded and sent as part of the request to Azure OpenAI.
This step is not mandatory; the model will still function correctly and without errors even if the prompts are not refined using this script. However, we strongly recommend using it to ensure the generation of high-quality videos.
### How to run:
Run the script for **text-to-video**:
$ python convert_demo.py --prompt "A girl riding a bike." --type "t2v"
Note:
Please set the OPENAI_API_KEY and OPENAI_BASE_URL(if needed) environment variable to your OpenAI API key before running this script.
Run the script:
$ python convert_demo.py --prompt "A girl ridding a bike." # Using with OpenAI's API
Run the script for **image-to-video**:
$ python convert_demo.py --prompt "the cat is running" --type "i2v" --image_path "/path/to/your/image.jpg"
"""
import argparse
from openai import OpenAI, AzureOpenAI
import base64
from mimetypes import guess_type
from openai import OpenAI
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.
sys_prompt_t2v = """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:
@ -29,19 +34,50 @@ Other times the user will not want modifications , but instead want a new image
Video descriptions must have the same num of words as examples below. Extra words will be ignored.
"""
sys_prompt_i2v = """
**Objective**: **Give a highly descriptive video caption based on input image and user input. **. As an expert, delve deep into the image with a discerning eye, leveraging rich creativity, meticulous thought. When describing the details of an image, include appropriate dynamic information to ensure that the video caption contains reasonable actions and plots. If user input is not empty, then the caption should be expanded according to the user's input.
def convert_prompt(prompt: str, retry_times: int = 3):
**Note**: The input image is the first frame of the video, and the output video caption should describe the motion starting from the current image. User input is optional and can be empty.
**Note**: Don't contain camera transitions!!! Don't contain screen switching!!! Don't contain perspective shifts !!!
**Answering Style**:
Answers should be comprehensive, conversational, and use complete sentences. The answer should be in English no matter what the user's input is. Provide context where necessary and maintain a certain tone. Begin directly without introductory phrases like "The image/video showcases" "The photo captures" and more. For example, say "A woman is on a beach", instead of "A woman is depicted in the image".
**Output Format**: "[highly descriptive image caption here]"
user input:
"""
def image_to_url(image_path):
mime_type, _ = guess_type(image_path)
if mime_type is None:
mime_type = "application/octet-stream"
with open(image_path, "rb") as image_file:
base64_encoded_data = base64.b64encode(image_file.read()).decode("utf-8")
return f"data:{mime_type};base64,{base64_encoded_data}"
def convert_prompt(prompt: str, retry_times: int = 3, type: str = "t2v", image_path: str = None):
"""
Convert a prompt to a format that can be used by the model for inference
"""
client = OpenAI()
text = prompt.strip()
## If you using with Azure OpenAI, please uncomment the below line and comment the above line
# client = AzureOpenAI(
# api_key="",
# api_version="",
# azure_endpoint=""
# )
text = prompt.strip()
for i in range(retry_times):
if type == "t2v":
response = client.chat.completions.create(
messages=[
{"role": "system", "content": f"{sys_prompt}"},
{"role": "system", "content": f"{sys_prompt_t2v}"},
{
"role": "user",
"content": 'Create an imaginative video descriptive caption or modify an earlier caption for the user input : " a girl is on the beach"',
@ -71,7 +107,30 @@ def convert_prompt(prompt: str, retry_times: int = 3):
"content": f'Create an imaginative video descriptive caption or modify an earlier caption in ENGLISH for the user input: " {text} "',
},
],
model="glm-4-0520", # glm-4-0520 and gpt-4o have be tested
model="glm-4-plus", # glm-4-plus and gpt-4o have be tested
temperature=0.01,
top_p=0.7,
stream=False,
max_tokens=250,
)
else:
response = client.chat.completions.create(
model="gpt-4o",
messages=[
{"role": "system", "content": f"{sys_prompt_i2v}"},
{
"role": "user",
"content": [
{"type": "text", "text": prompt},
{
"type": "image_url",
"image_url": {
"url": image_to_url(image_path),
},
},
],
},
],
temperature=0.01,
top_p=0.7,
stream=False,
@ -86,7 +145,9 @@ if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--prompt", type=str, required=True, help="Prompt to convert")
parser.add_argument("--retry_times", type=int, default=3, help="Number of times to retry the conversion")
parser.add_argument("--type", type=str, default="t2v", help="Type of conversion (t2v or i2v)")
parser.add_argument("--image_path", type=str, default=None, help="Path to the image file")
args = parser.parse_args()
converted_prompt = convert_prompt(args.prompt, args.retry_times)
converted_prompt = convert_prompt(args.prompt, args.retry_times, args.type, args.image_path)
print(converted_prompt)

21
sat/diffusion_video.py
View File

@ -162,7 +162,6 @@ class SATVideoDiffusionEngine(nn.Module):
kwargs = {"timesteps": len(z[n * n_samples : (n + 1) * n_samples])}
else:
kwargs = {}
use_cp = False
out = self.first_stage_model.decode(z[n * n_samples : (n + 1) * n_samples], **kwargs)
all_out.append(out)
out = torch.cat(all_out, dim=0)
@ -176,8 +175,6 @@ class SATVideoDiffusionEngine(nn.Module):
x = x.permute(0, 2, 1, 3, 4).contiguous()
return x * self.scale_factor # already encoded
use_cp = False
n_samples = default(self.en_and_decode_n_samples_a_time, x.shape[0])
n_rounds = math.ceil(x.shape[0] / n_samples)
all_out = []
@ -305,6 +302,24 @@ class SATVideoDiffusionEngine(nn.Module):
samples = self.sample(c, shape=z.shape[1:], uc=uc, batch_size=N, **sampling_kwargs) # b t c h w
samples = samples.permute(0, 2, 1, 3, 4).contiguous()
if self.noised_image_input:
image = x[:, :, 0:1]
image = self.add_noise_to_first_frame(image)
image = self.encode_first_stage(image, batch)
image = image.permute(0, 2, 1, 3, 4).contiguous()
image = torch.concat([image, torch.zeros_like(z[:, 1:])], dim=1)
c["concat"] = image
uc["concat"] = image
samples = self.sample(c, shape=z.shape[1:], uc=uc, batch_size=N, **sampling_kwargs) # b t c h w
samples = samples.permute(0, 2, 1, 3, 4).contiguous()
if only_log_video_latents:
latents = 1.0 / self.scale_factor * samples
log["latents"] = latents
else:
samples = self.decode_first_stage(samples).to(torch.float32)
samples = samples.permute(0, 2, 1, 3, 4).contiguous()
log["samples"] = samples
else:
if only_log_video_latents:
latents = 1.0 / self.scale_factor * samples
log["latents"] = latents

41
sat/dit_video_concat.py
View File

@ -155,6 +155,25 @@ def get_1d_sincos_pos_embed_from_grid(embed_dim, pos):
return emb
class Basic2DPositionEmbeddingMixin(BaseMixin):
def __init__(self, height, width, compressed_num_frames, hidden_size, text_length=0):
super().__init__()
self.height = height
self.width = width
self.spatial_length = height * width
self.pos_embedding = nn.Parameter(
torch.zeros(1, int(text_length + self.spatial_length), int(hidden_size)), requires_grad=False
)
def position_embedding_forward(self, position_ids, **kwargs):
return self.pos_embedding
def reinit(self, parent_model=None):
del self.transformer.position_embeddings
pos_embed = get_2d_sincos_pos_embed(self.pos_embedding.shape[-1], self.height, self.width)
self.pos_embedding.data[:, -self.spatial_length :].copy_(torch.from_numpy(pos_embed).float().unsqueeze(0))
class Basic3DPositionEmbeddingMixin(BaseMixin):
def __init__(
self,
@ -240,10 +259,10 @@ class Rotary3DPositionEmbeddingMixin(BaseMixin):
text_length,
theta=10000,
rot_v=False,
learnable_pos_embed=False,
):
super().__init__()
self.rot_v = rot_v
self.text_length = text_length
dim_t = hidden_size_head // 4
dim_h = hidden_size_head // 8 * 3
@ -274,6 +293,13 @@ class Rotary3DPositionEmbeddingMixin(BaseMixin):
self.register_buffer("freqs_sin", freqs_sin)
self.register_buffer("freqs_cos", freqs_cos)
self.text_length = text_length
if learnable_pos_embed:
num_patches = height * width * compressed_num_frames + text_length
self.pos_embedding = nn.Parameter(torch.zeros(1, num_patches, int(hidden_size)), requires_grad=True)
else:
self.pos_embedding = None
def rotary(self, t, **kwargs):
seq_len = t.shape[2]
freqs_cos = self.freqs_cos[:seq_len].unsqueeze(0).unsqueeze(0)
@ -362,7 +388,6 @@ class FinalLayerMixin(BaseMixin):
def final_forward(self, logits, **kwargs):
x, emb = logits[:, kwargs["text_length"] :, :], kwargs["emb"] # x:(b,(t n),d)
shift, scale = self.adaLN_modulation(emb).chunk(2, dim=1)
x = modulate(self.norm_final(x), shift, scale)
x = self.linear(x)
@ -458,6 +483,7 @@ class AdaLNMixin(BaseMixin):
# hidden_states (b,(n_t+t*n_i),d)
text_hidden_states = hidden_states[:, :text_length] # (b,n,d)
img_hidden_states = hidden_states[:, text_length:] # (b,(t n),d)
layer = self.transformer.layers[kwargs["layer_id"]]
adaLN_modulation = self.adaLN_modulations[kwargs["layer_id"]]
@ -492,7 +518,6 @@ class AdaLNMixin(BaseMixin):
attention_output = layer.attention(attention_input, mask, **kwargs)
text_attention_output = attention_output[:, :text_length] # (b,n,d)
img_attention_output = attention_output[:, text_length:] # (b,(t n),d)
if self.transformer.layernorm_order == "sandwich":
text_attention_output = layer.third_layernorm(text_attention_output)
img_attention_output = layer.third_layernorm(img_attention_output)
@ -748,6 +773,15 @@ class DiffusionTransformer(BaseModel):
b, t, d, h, w = x.shape
if x.dtype != self.dtype:
x = x.to(self.dtype)
# This is not use in inference
if "concat_images" in kwargs and kwargs["concat_images"] is not None:
if kwargs["concat_images"].shape[0] != x.shape[0]:
concat_images = kwargs["concat_images"].repeat(2, 1, 1, 1, 1)
else:
concat_images = kwargs["concat_images"]
x = torch.cat([x, concat_images], dim=2)
assert (y is not None) == (
self.num_classes is not None
), "must specify y if and only if the model is class-conditional"
@ -768,5 +802,4 @@ class DiffusionTransformer(BaseModel):
kwargs["input_ids"] = kwargs["position_ids"] = kwargs["attention_mask"] = torch.ones((1, 1)).to(x.dtype)
output = super().forward(**kwargs)[0]
return output

31
sat/sample_video.py
View File

@ -11,6 +11,7 @@ import numpy as np
from einops import rearrange
import torchvision.transforms as TT
from sat.model.base_model import get_model
from sat.training.model_io import load_checkpoint
from sat import mpu
@ -19,6 +20,7 @@ from diffusion_video import SATVideoDiffusionEngine
from arguments import get_args
from torchvision.transforms.functional import center_crop, resize
from torchvision.transforms import InterpolationMode
from PIL import Image
def read_from_cli():
@ -132,6 +134,11 @@ def sampling_main(args, model_cls):
image_size = [480, 720]
if args.image2video:
chained_trainsforms = []
chained_trainsforms.append(TT.ToTensor())
transform = TT.Compose(chained_trainsforms)
sample_func = model.sample
T, H, W, C, F = args.sampling_num_frames, image_size[0], image_size[1], args.latent_channels, 8
num_samples = [1]
@ -139,10 +146,21 @@ def sampling_main(args, model_cls):
device = model.device
with torch.no_grad():
for text, cnt in tqdm(data_iter):
# reload model on GPU
model.to(device)
print("rank:", rank, "start to process", text, cnt)
# TODO: broadcast image2video
if args.image2video:
text, image_path = text.split("@@")
assert os.path.exists(image_path), image_path
image = Image.open(image_path).convert("RGB")
image = transform(image).unsqueeze(0).to("cuda")
image = resize_for_rectangle_crop(image, image_size, reshape_mode="center").unsqueeze(0)
image = image * 2.0 - 1.0
image = image.unsqueeze(2).to(torch.bfloat16)
image = model.encode_first_stage(image, None)
image = image.permute(0, 2, 1, 3, 4).contiguous()
pad_shape = (image.shape[0], T - 1, C, H // F, W // F)
image = torch.concat([image, torch.zeros(pad_shape).to(image.device).to(image.dtype)], dim=1)
else:
image = None
value_dict = {
"prompt": text,
"negative_prompt": "",
@ -168,6 +186,11 @@ def sampling_main(args, model_cls):
for k in c:
if not k == "crossattn":
c[k], uc[k] = map(lambda y: y[k][: math.prod(num_samples)].to("cuda"), (c, uc))
if args.image2video and image is not None:
c["concat"] = image
uc["concat"] = image
for index in range(args.batch_size):
# reload model on GPU
model.to(device)

11
sat/sgm/modules/diffusionmodules/loss.py
View File

@ -2,17 +2,9 @@ from typing import List, Optional, Union
import torch
import torch.nn as nn
import torch.nn.functional as F
from omegaconf import ListConfig
import math
from ...modules.diffusionmodules.sampling import VideoDDIMSampler, VPSDEDPMPP2MSampler
from ...util import append_dims, instantiate_from_config
from ...modules.autoencoding.lpips.loss.lpips import LPIPS
# import rearrange
from einops import rearrange
import random
from sat import mpu
@ -107,6 +99,9 @@ class VideoDiffusionLoss(StandardDiffusionLoss):
(1 - alphas_cumprod_sqrt**2) ** 0.5, input.ndim
)
if "concat_images" in batch.keys():
additional_model_inputs["concat_images"] = batch["concat_images"]
model_output = denoiser(network, noised_input, alphas_cumprod_sqrt, cond, **additional_model_inputs)
w = append_dims(1 / (1 - alphas_cumprod_sqrt**2), input.ndim) # v-pred

4
sat/vae_modules/cp_enc_dec.py
View File

@ -807,7 +807,7 @@ class ContextParallelEncoder3D(nn.Module):
kernel_size=3,
)
def forward(self, x):
def forward(self, x, **kwargs):
# timestep embedding
temb = None
@ -948,7 +948,7 @@ class ContextParallelDecoder3D(nn.Module):
kernel_size=3,
)
def forward(self, z, clear_fake_cp_cache=True):
def forward(self, z, clear_fake_cp_cache=True, **kwargs):
self.last_z_shape = z.shape
# timestep embedding