[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"project-72284":3},{"id":4,"name":5,"fullName":6,"owner":7,"repo":5,"description":8,"homepage":9,"htmlUrl":10,"language":11,"languages":10,"totalLinesOfCode":10,"stars":12,"forks":13,"watchers":14,"openIssues":15,"contributorsCount":16,"subscribersCount":16,"size":16,"stars1d":17,"stars7d":18,"stars30d":19,"stars90d":16,"forks30d":16,"starsTrendScore":20,"compositeScore":21,"rankGlobal":10,"rankLanguage":10,"license":22,"archived":23,"fork":23,"defaultBranch":24,"hasWiki":25,"hasPages":25,"topics":26,"createdAt":10,"pushedAt":10,"updatedAt":33,"readmeContent":34,"aiSummary":35,"trendingCount":16,"starSnapshotCount":16,"syncStatus":36,"lastSyncTime":37,"discoverSource":38},72284,"FastVideo","hao-ai-lab\u002FFastVideo","hao-ai-lab","A unified inference and post-training framework for accelerated video generation.","https:\u002F\u002Fhao-ai-lab.github.io\u002FFastVideo",null,"Python",3702,359,37,53,0,17,44,238,51,29.67,"Apache License 2.0",false,"main",true,[27,28,29,30,31,32],"diffusers","diffusion-models","distillation","inference","post-training","video-generation","2026-06-12 02:03:01","\u003Cdiv align=\"center\">\n\u003Cimg src=assets\u002Flogos\u002Flogo.svg width=\"30%\"\u002F>\n\u003C\u002Fdiv>\n\n\u003Cp align=\"center\">\n     | \u003Ca href=\"https:\u002F\u002Fhao-ai-lab.github.io\u002FFastVideo\">\u003Cb>Documentation\u003C\u002Fb>\u003C\u002Fa> | \u003Ca href=\"https:\u002F\u002Fhao-ai-lab.github.io\u002FFastVideo\u002Finference\u002Finference_quick_start\u002F\">\u003Cb> Quick Start\u003C\u002Fb>\u003C\u002Fa> | \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fhao-ai-lab\u002FFastVideo\u002Fdiscussions\u002F982\"  target=\"_blank\">\u003Cb>Weekly Dev Meeting\u003C\u002Fb>\u003C\u002Fa>  | 🟣💬 \u003Ca href=\"https:\u002F\u002Fjoin.slack.com\u002Ft\u002Ffastvideo\u002Fshared_invite\u002Fzt-3f4lao1uq-u~Ipx6Lt4J27AlD2y~IdLQ\" target=\"_blank\"> \u003Cb>Slack\u003C\u002Fb> \u003C\u002Fa> |  🟣💬 \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fhao-ai-lab\u002FFastVideo\u002Fdiscussions\u002F1097\" target=\"_blank\"> \u003Cb> WeChat \u003C\u002Fb> \u003C\u002Fa> |\n\u003C\u002Fp>\n\n**FastVideo is a unified post-training and real-time inference framework for accelerated video generation.**\n\n## NEWS\n- `2026\u002F03\u002F17`: Release Live demo: [Into the Dreamverse: Vibe Directing in FastVideo](https:\u002F\u002Fdreamverse.fastvideo.org\u002F), check out the [Blog](https:\u002F\u002Fhaoailab.com\u002Fblogs\u002Fdreamverse\u002F).\n- `2026\u002F03\u002F13`: Release Live demo: [Create a 5s 1080p Video in 4.5s with FastVideo on a Single GPU](https:\u002F\u002F1080p.fastvideo.org\u002F), check out the [Blog](https:\u002F\u002Fhaoailab.com\u002Fblogs\u002Ffastvideo_realtime_1080p\u002F).\n- `2025\u002F11\u002F19`: Release [CausalWan2.2 I2V A14B Preview](https:\u002F\u002Fhuggingface.co\u002FFastVideo\u002FCausalWan2.2-I2V-A14B-Preview-Diffusers) models, [Blog](https:\u002F\u002Fhao-ai-lab.github.io\u002Fblogs\u002Ffastvideo_causalwan_preview\u002F) and [Inference Code!](https:\u002F\u002Fgithub.com\u002Fhao-ai-lab\u002FFastVideo\u002Fblob\u002Fmain\u002Fexamples\u002Finference\u002Fbasic\u002Fbasic_self_forcing_causal_wan2_2_i2v.py).\n- `2025\u002F08\u002F04`: Release [FastWan](https:\u002F\u002Fhao-ai-lab.github.io\u002FFastVideo\u002Fdistillation\u002Fdmd) models and [Sparse-Distillation](https:\u002F\u002Fhao-ai-lab.github.io\u002Fblogs\u002Ffastvideo_post_training\u002F).\n\n### More News\n\n- `2025\u002F06\u002F14`: Release finetuning and inference code for [VSA](https:\u002F\u002Farxiv.org\u002Fpdf\u002F2505.13389).\n- `2025\u002F04\u002F24`: [FastVideo V1](https:\u002F\u002Fhao-ai-lab.github.io\u002Fblogs\u002Ffastvideo\u002F) is released!\n- `2025\u002F02\u002F18`: Release the inference code for [Sliding Tile Attention](https:\u002F\u002Fhao-ai-lab.github.io\u002Fblogs\u002Fsta\u002F).\n\n## Key Features\n\nFastVideo has the following features:\n\n- End-to-end post-training support for bidirectional and autoregressive models:\n  - Support full finetuning and LoRA finetuning for state-of-the-art open video DiTs\n  - Data preprocessing pipeline for video, image, and text data\n  - Distribution Matching Distillation (DMD2) stepwise distillation.\n  - Sparse attention with [Video Sparse Attention](https:\u002F\u002Farxiv.org\u002Fpdf\u002F2505.13389)\n  - [Sparse distillation](https:\u002F\u002Fhao-ai-lab.github.io\u002Fblogs\u002Ffastvideo_post_training\u002F) to achieve >50x denoising speedup\n  - Scalable training with FSDP2, sequence parallelism, and selective activation checkpointing.\n  - Causal distillation through Self-Forcing\n  - See this [page](https:\u002F\u002Fhao-ai-lab.github.io\u002FFastVideo\u002Ftraining\u002Foverview\u002F) for full list of supported models and recipes.\n- State-of-the-art performance optimizations for inference\n  - Sequence Parallelism for distributed inference\n  - Multiple state-of-the-art attention backends\n  - User-friendly CLI and Python API\n  - See this [page](https:\u002F\u002Fhao-ai-lab.github.io\u002FFastVideo\u002Finference\u002Foptimizations\u002F) for full list of supported optimizations.\n- Diverse hardware and OS support\n  - Support H100, A100, 4090\n  - Support Linux, Windows, MacOS\n  - See this [page](https:\u002F\u002Fhao-ai-lab.github.io\u002FFastVideo\u002Finference\u002Fsupport_matrix\u002F) for full list of supported models, hardware assumptions, and optimization compatibility.\n\n## Getting Started\n\nWe recommend using [uv](https:\u002F\u002Fdocs.astral.sh\u002Fuv\u002F) to create a clean environment. If you previously used Conda, switching to uv generally gives faster and more stable installs.\n\n```bash\n# Create and activate a new uv environment\nuv venv --python 3.12 --seed\nsource .venv\u002Fbin\u002Factivate\n\n# Install FastVideo\nuv pip install fastvideo\n```\n\nPlease see our [docs](https:\u002F\u002Fhao-ai-lab.github.io\u002FFastVideo\u002Fgetting_started\u002Finstallation\u002F) for more detailed installation instructions.\n\n## Sparse Distillation\n\nFor our sparse distillation techniques, please see our [distillation docs](https:\u002F\u002Fhao-ai-lab.github.io\u002FFastVideo\u002Fdistillation\u002Fdmd\u002F) and check out our [blog](https:\u002F\u002Fhao-ai-lab.github.io\u002Fblogs\u002Ffastvideo_post_training\u002F).\n\nSee below for recipes and datasets:\n\n| Model                                                                                 | Sparse Distillation                                                                                             | Dataset                                                                                                  |\n| ------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------- |\n| [FastWan2.1-T2V-1.3B](https:\u002F\u002Fhuggingface.co\u002FFastVideo\u002FFastWan2.1-T2V-1.3B-Diffusers) | [Recipe](https:\u002F\u002Fgithub.com\u002Fhao-ai-lab\u002FFastVideo\u002Ftree\u002Fmain\u002Fexamples\u002Fdistill\u002FWan2.1-T2V\u002FWan-Syn-Data-480P)       | [FastVideo Synthetic Wan2.1 480P](https:\u002F\u002Fhuggingface.co\u002Fdatasets\u002FFastVideo\u002FWan-Syn_77x448x832_600k)     |\n| [FastWan2.2-TI2V-5B](https:\u002F\u002Fhuggingface.co\u002FFastVideo\u002FFastWan2.2-TI2V-5B-Diffusers)   | [Recipe](https:\u002F\u002Fgithub.com\u002Fhao-ai-lab\u002FFastVideo\u002Ftree\u002Fmain\u002Fexamples\u002Fdistill\u002FWan2.2-TI2V-5B-Diffusers\u002FData-free) | [FastVideo Synthetic Wan2.2 720P](https:\u002F\u002Fhuggingface.co\u002Fdatasets\u002FFastVideo\u002FWan2.2-Syn-121x704x1280_32k) |\n\n## Inference\n\n### Generating Your First Video\n\nHere's a minimal example to generate a video using the default settings. Make sure VSA kernels are [installed](https:\u002F\u002Fhao-ai-lab.github.io\u002FFastVideo\u002Fattention\u002Fvsa\u002F#installation). Create a file called `example.py` with the following code:\n\n```python\nimport os\nfrom fastvideo import VideoGenerator\n\ndef main():\n    os.environ[\"FASTVIDEO_ATTENTION_BACKEND\"] = \"VIDEO_SPARSE_ATTN\"\n\n    # Create a video generator with a pre-trained model\n    generator = VideoGenerator.from_pretrained(\n        \"FastVideo\u002FFastWan2.1-T2V-1.3B-Diffusers\",\n        num_gpus=1,  # Adjust based on your hardware\n    )\n\n    # Define a prompt for your video\n    prompt = \"A curious raccoon peers through a vibrant field of yellow sunflowers, its eyes wide with interest.\"\n\n    # Generate the video\n    video = generator.generate_video(\n        prompt,\n        output_path=\"my_videos\u002F\",  # Controls where videos are saved\n        save_video=True\n    )\n\nif __name__ == '__main__':\n    main()\n```\n\nRun the script with:\n\n```bash\npython example.py\n```\n\nFor a more detailed guide, please see our [inference quick start](https:\u002F\u002Fhao-ai-lab.github.io\u002FFastVideo\u002Finference\u002Finference_quick_start\u002F).\n\n## More Guides\n\n- [Design Overview](https:\u002F\u002Fhao-ai-lab.github.io\u002FFastVideo\u002Fdesign\u002Foverview\u002F)\n- [Distillation Guide](https:\u002F\u002Fhao-ai-lab.github.io\u002FFastVideo\u002Fdistillation\u002Fdmd\u002F)\n- [Contribution Guide](https:\u002F\u002Fhao-ai-lab.github.io\u002FFastVideo\u002Fcontributing\u002Foverview\u002F)\n\n## Awesome work using FastVideo or our research projects\n\n- [SGLang](https:\u002F\u002Fgithub.com\u002Fsgl-project\u002Fsglang\u002Ftree\u002Fmain\u002Fpython\u002Fsglang\u002Fmultimodal_gen): SGLang's diffusion inference functionality is based on a fork of FastVideo on Sept. 24, 2025.\n- [DanceGRPO](https:\u002F\u002Fgithub.com\u002FXueZeyue\u002FDanceGRPO): A unified framework to adapt Group Relative Policy Optimization (GRPO) to visual generation paradigms. Code based on FastVideo.\n- [SRPO](https:\u002F\u002Fgithub.com\u002FTencent-Hunyuan\u002FSRPO): A method to directly align the full diffusion trajectory with fine-grained human preference. Code based on FastVideo.\n- [DCM](https:\u002F\u002Fgithub.com\u002FVchitect\u002FDCM): Dual-expert consistency model for efficient and high-quality video generation. Code based on FastVideo.\n- [HY-WorldPlay](https:\u002F\u002Fgithub.com\u002FTencent-Hunyuan\u002FHY-WorldPlay): An action-conditioned world model model trained using FastVideo framework.\n- [Hunyuan Video 1.5](https:\u002F\u002Fgithub.com\u002FTencent-Hunyuan\u002FHunyuanVideo-1.5): A leading lightweight video generation model, where they proposed SSTA based on Sliding Tile Attention.\n- [Kandinsky-5.0](https:\u002F\u002Fgithub.com\u002Fkandinskylab\u002Fkandinsky-5): A family of diffusion models for video & image generation, where their NABLA attention includes a Sliding Tile Attention branch.\n- [LongCat Video](https:\u002F\u002Fgithub.com\u002Fmeituan-longcat\u002FLongCat-Video): A foundational video generation model with 13.6B parameters with block-sparse attention similar to Video Sparse Attention.\n\n## 🤝 Contributing\n\nWe welcome all contributions. Please check out our guide [here](https:\u002F\u002Fhao-ai-lab.github.io\u002FFastVideo\u002Fcontributing\u002Foverview\u002F).\nSee details in [development roadmap](https:\u002F\u002Fgithub.com\u002Fhao-ai-lab\u002FFastVideo\u002Fissues\u002F899).\n\n## Acknowledgement\n\nWe learned the design and reused code from the following projects: [Wan-Video](https:\u002F\u002Fgithub.com\u002FWan-Video), [ThunderKittens](https:\u002F\u002Fgithub.com\u002FHazyResearch\u002FThunderKittens), [DMD2](https:\u002F\u002Fgithub.com\u002Ftianweiy\u002FDMD2), [diffusers](https:\u002F\u002Fgithub.com\u002Fhuggingface\u002Fdiffusers), [xDiT](https:\u002F\u002Fgithub.com\u002Fxdit-project\u002FxDiT), [vLLM](https:\u002F\u002Fgithub.com\u002Fvllm-project\u002Fvllm), [SGLang](https:\u002F\u002Fgithub.com\u002Fsgl-project\u002Fsglang). We thank [MBZUAI](https:\u002F\u002Fifm.mbzuai.ac.ae\u002F), [Anyscale](https:\u002F\u002Fwww.anyscale.com\u002F), and [GMI Cloud](https:\u002F\u002Fwww.gmicloud.ai\u002F) for their support throughout this project.\n\n## Citation\n\nIf you find FastVideo useful, please consider citing our research work:\n\n```bibtex\n@article{zhang2025vsa,\n  title={Vsa: Faster video diffusion with trainable sparse attention},\n  author={Zhang, Peiyuan and Chen, Yongqi and Huang, Haofeng and Lin, Will and Liu, Zhengzhong and Stoica, Ion and Xing, Eric and Zhang, Hao},\n  journal={arXiv preprint arXiv:2505.13389},\n  year={2025}\n}\n\n@article{zhang2025fast,\n  title={Fast video generation with sliding tile attention},\n  author={Zhang, Peiyuan and Chen, Yongqi and Su, Runlong and Ding, Hangliang and Stoica, Ion and Liu, Zhengzhong and Zhang, Hao},\n  journal={arXiv preprint arXiv:2502.04507},\n  year={2025}\n}\n```\n","FastVideo 是一个用于加速视频生成的统一后训练和实时推理框架。它支持双向及自回归模型的端到端后训练，包括全微调、LoRA 微调以及数据预处理管道，并采用分布匹配蒸馏（DMD2）和稀疏注意力等技术来实现超过50倍的去噪速度提升。此外，FastVideo 还通过序列并行化、FSDP2 和选择性激活检查点等方法优化了大规模训练过程。该项目特别适合需要高效视频生成的应用场景，如创意内容制作、虚拟现实体验开发等领域。",2,"2026-06-11 03:41:12","high_star"]