mirror of
https://github.com/THUDM/CogVideo.git
synced 2026-05-06 22:58:13 +08:00
2d3f2a4d02
Implements a comprehensive batch video generation tool that addresses the #1 missing feature for production users: generating multiple videos from a single batch file instead of one-at-a-time processing. ## New Files ### tools/batch_inference.py Production-ready batch inference script with: **Core Features:** - JSONL input format (one job per line, streaming-friendly) - Support for all generation types: t2v, i2v, v2v - Progress tracking with tqdm (progress bar, ETA) - Robust error handling (logs errors, continues batch) - Resume capability (tracks completed jobs, skips on restart) **Input Schema:** - prompt (required): Text description - output_name (required): Output filename - image_path (optional): For i2v generation - video_path (optional): For v2v generation - num_frames, guidance_scale, num_inference_steps, seed, width, height (optional) **Multi-GPU Support:** - Job-level parallelism via --gpu_id and --num_gpus flags - Each GPU processes a subset of jobs (round-robin distribution) - State file prevents duplicate work across processes **Memory Management:** - Loads model once, generates sequentially - CPU offloading enabled by default - VAE slicing and tiling enabled ### resources/example_batch_*.jsonl Example batch files demonstrating: - example_batch_t2v.jsonl: Text-to-video prompts - example_batch_i2v.jsonl: Image-to-video with image_path - example_batch_v2v.jsonl: Video-to-video with video_path ## Design Decisions 1. **JSONL over JSON**: Better for large batches, streaming, and manual editing 2. **Reuse generation logic**: Mirrors cli_demo.py patterns for consistency 3. **Single model per batch**: Memory efficient, simpler implementation 4. **State persistence**: JSON state file enables reliable resume 5. **Error isolation**: One failed job doesn't stop the batch ## Usage Examples # Basic text-to-video python tools/batch_inference.py --batch_file prompts.jsonl --model_path THUDM/CogVideoX1.5-5B # Multi-GPU (4 GPUs) for i in {0..3}; do CUDA_VISIBLE_DEVICES=$i python tools/batch_inference.py --batch_file batch.jsonl --gpu_id $i --num_gpus 4 & done
9 lines
966 B
JSON
9 lines
966 B
JSON
# Example video-to-video batch file
|
|
# For v2v generation, include video_path pointing to your input videos
|
|
# Run with: python tools/batch_inference.py --batch_file resources/example_batch_v2v.jsonl --model_path THUDM/CogVideoX1.5-5B --generate_type v2v
|
|
|
|
{"prompt": "Transform this video into a watercolor painting style with soft brushstrokes", "output_name": "watercolor_style.mp4", "video_path": "./input_videos/original1.mp4"}
|
|
{"prompt": "Convert to anime style with vibrant colors and dramatic lighting", "output_name": "anime_style.mp4", "video_path": "./input_videos/original2.mp4", "guidance_scale": 7.5}
|
|
{"prompt": "Add cinematic color grading with film grain and dramatic contrast", "output_name": "cinematic_grade.mp4", "video_path": "./input_videos/original3.mp4", "num_inference_steps": 40}
|
|
{"prompt": "Transform into a vintage black and white film with classic aesthetics", "output_name": "vintage_bw.mp4", "video_path": "./input_videos/original4.mp4"}
|