A few-step autoregressive video distillation method by the ZIP Lab at Zhejiang University and others, addressing the issue where subjects or scenes "change appearance" in long videos after leaving the frame and returning. The approach involves having the teacher predict the same noisy student block under two contexts—differing only by an informative history (earlier, less drifted student blocks, or teacher rollouts from a fixed viewpoint that keeps the scene in view). The difference between these predictions is used as a residual, which is weighted and added to the score/velocity target for distillation, after which the student aligns with this corrected target. The same framework is applied separately to 60-second long video generation and world models for "off-screen scene evolution."
Likes 5 · Code Stars 27
Reasons
- The authors report that for the 1.3B model, the VBench quality score for 15-second videos is 85.07, and the overall VBench-Long score for 60-second videos is 79.82, claiming these are the best among models of the same scale. For off-screen world modeling, PlayWorld evolution quality improved from 2.35 to 2.70, and MemoBench object consistency improved from 58.43 to 62.18.
- Comprehensive availability: The long video generator (based on the Wan2.1-T2V-1.3B base), the world model (based on the DreamX-World-5B base), and the AR teacher checkpoints, along with training data and PlayWorld / MemoBench evaluation inputs, are all on Hugging Face. Inference, training, and evaluation code are in the repository, with code licensed under Apache-2.0.
- Like the top recommendation, this follows the Wan2.1-1.3B few-step autoregressive path; one addresses speed while the other addresses long-term consistency, making them suitable for comparative reading.
Usage
- Environment: Linux + NVIDIA GPU (CUDA 12.8 compatible), create a Python 3.12 environment with conda, install torch 2.8.0, then run pip install -r requirements.txt.
- Long videos: Run hf download Wan-AI/Wan2.1-T2V-1.3B --local-dir models/Wan2.1-T2V-1.3B and hf download xzxxntxdy/History-Forcing dmd/model.pt --local-dir models/History-Forcing, then run bash scripts/long_video/inference.sh --prompt "…" --output-dir output/demo/long_video to generate 60-second videos.
- World model: Download Wan2.2-TI2V-5B, GD-ML/DreamX-World-5B, and opsd/model.pt, along with the dataset, then run bash scripts/world_model/run.sh inference --checkpoint models/History-Forcing/opsd/model.pt --num-samples 1 (inputs include the first frame image, scene prompt, and camera actions).
Risks
- The README does not specify VRAM requirements or generation speed; training requires downloading additional Causal-Forcing initialization weights and Self-Forcing prompts, and uses W&B for logging.
- Checkpoints are based on Wan2.1 / Wan2.2 and DreamX-World; third-party components have their own licenses, and the model card does not separately specify a license.
- Scores are reported by the authors; the scoring method for PlayWorld evolution quality requires consulting the paper's main text. Since the generated videos have not been viewed, no assessment of visual quality is provided.