Top AI Repos — open-source AI, indexed and scored
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
(CVPR 2025) From Slow Bidirectional to Fast Autoregressive Video Diffusion Models
| Date | Stars |
|---|---|
| 2026-07-31 | 1410 |
| 2026-08-04 | 1412 |
| 2026-08-06 | 1412 |
Today
— stars today
This week
— stars this week
This month
— stars this month
Momentum
0.0
growth rate 0.00%/day
# From Slow Bidirectional to Fast Autoregressive Video Diffusion Models [[Huggingface](https://huggingface.co/tianweiy/CausVid)][[Project](https://causvid.github.io/)] Few-step Text-to-Video Generation.  > [**From Slow Bidirectional to Fast Autoregressive Video Diffusion Models**](https://causvid.github.io/), > Tianwei Yin*, Qiang Zhang*, Richard Zhang, William T. Freeman, Frédo Durand, Eli Shechtman, Xun Huang (* equal contribution) > *CVPR 2025 ([arXiv 2412.07772](https://arxiv.org/abs/2412.07772))* ## Abstract Current video diffusion models achieve impressive generation quality but struggle in interactive applications due to bidirectional attention dependencies. The generation of a single frame requires the model to process the entire sequence, including the future. We address this limitation by adapting a pretrained bidirectional diffusion transformer to an autoregressive transformer that generates frames on-the-fly. To further reduce latency, we extend distribution matching distillation (DMD) to videos, distilling 50-step diffusion model into a 4-step generator. To enable stable and high-quality distillation, we introduce a student initialization scheme based on teacher's ODE trajectories, as well as an asymmetric distillation strategy that supervises a causal student model with a bidirectional teacher. This approach effectively mitigates error accumulation in autoregressive generation, allowing long-duration video synthesis despite training on short clips. Our model achieves a total score of 84.27 on the VBench-Long benchmark, surpassing all previous video generation models. It enables fast streaming generation of high-quality videos at 9.4 FPS on a single GPU thanks to KV caching. Our approach also enables streaming video-to-video translation, image-to-video, and dynamic prompting in a zero-shot manner. <div style="border-top: 2px solid red; padding-top: 10px; margin-top: 10px; margin-bottom: 10px;"> ⚠️ This repo is a work in progress. Expect frequent updates in the coming weeks. </div> ## Environment Setup ```bash conda create -n causvid python=3.10 -y conda activate causvid pip install torch torchvision pip install -r requirements.txt python setup.py develop ``` Also download the Wan base models from [here](https://github.com/Wan-Video/Wan2.1) and save it to wan_models/Wan2.1-T2V-1.3B/ ## Inference Example First download the checkpoints: [Autoregressive Model](https://huggingface.co/tianweiy/CausVid/tree/main/autoregressive_checkpoint), [Bidirectional Model 1](https://huggingface.co/tianweiy/CausVid/tree/main/bidirectional_checkpoint1) or [Bidirectional Model 2](https://huggingface.co/tianweiy/CausVid/tree/main/bidirectional_checkpoint2) (performs slightly better). ### Autoregressive 3-step 5-second Video Generation ```bash python minimal_inference/autoregressive_inference.py --config_path configs/wan_causal_dmd.yaml --checkpoint_folder XXX --output_folder XXX --prompt_file_path XXX ``` ### Autoregressive 3-step long Video Generation ```bash python minimal_inference/longvideo_autoregressive_inference.py --config_path configs/wan_causal_dmd.yaml --checkpoint_folder XXX --output_folder XXX --prompt_file_path XXX --num_rollout XXX ``` ### Bidirectional 3-step 5-second Video Generation ```bash python minimal_inference/bidirectional_inference.py --config_path configs/wan_bidirectional_dmd_from_scratch.yaml --checkpoint_folder XXX --output_folder XXX --prompt_file_path XXX ``` ## Training and Evaluation ### Dataset Preparation We use the [MixKit Dataset](https://huggingface.co/datasets/LanguageBind/Open-Sora-Plan-v1.1.0/tree/main/all_mixkit) (6K videos) as a toy example for distillation. To prepare the dataset, follow these steps. You can also download the final LMDB dataset from [here](https://huggingface.co/tianweiy/CausVid/tree/main/mixkit_latents_lmdb) ```bash # download and extract video from the Mixkit dataset pyt
Excerpt of 7,616 characters
Read on GitHub
8
Would you bet a product on this? Bounded 0–100 and slow moving.
matched fp:bddf129298580a4c, llm:description: "(CVPR 2025) From Slow Bidirectional to Fast Autoregressive Video Diffusion Models"; language: Python
matched fp:bddf129298580a4c, llm:description: "(CVPR 2025) From Slow Bidirectional to Fast Autoregressive Video Diffusion Models"; language: Python
matched fp:bddf129298580a4c, llm:description: "(CVPR 2025) From Slow Bidirectional to Fast Autoregressive Video Diffusion Models"; language: Python
matched fp:bddf129298580a4c, llm:description: "(CVPR 2025) From Slow Bidirectional to Fast Autoregressive Video Diffusion Models"; language: Python