FastVideo is a unified inference and post-training framework for accelerated video generation, hosted in the hao-ai-lab/FastVideo repository on GitHub. The project is positioned to cover both post-training and real-time inference, with the goal of achieving acceleration in video generation tasks. The project provides resource entry points such as documentation, a Cookbook, and a quick start guide.
-
Main capabilities Post-training support: The project provides end-to-end post-training support for bidirectional models and autoregressive models, specifically including: full fine-tuning and LoRA fine-tuning of current state-of-the-art open-source video DiT (Diffusion Transformer) models; data preprocessing pipelines for video, image, and text data; Distribution Matching Distillation (DMD2) progressive distillation; sparse attention using Video Sparse Attention; sparse distillation to achieve more than 50x denoising acceleration; scalable training through FSDP2, sequence parallelism, and selective activation checkpointing; causal distillation through Self-Forcing. The supported training workflows and model support matrix can be found on the project's dedicated pages. Inference optimization: Provides sequence parallelism (for distributed inference), multiple current state-of-the-art attention backends, and an easy-to-use CLI and Python API. Hardware and operating systems: Supports GPUs such as H100, A100, and 4090; at the operating system level, supports Linux, Windows, and macOS. The project notes that dedicated pages can be consulted for supported models, hardware assumptions, and optimization compatibility. Real-time video generation and editing: The project includes the Dreamverse platform, which can perform real-time generation and editing ("vibe directing") on video being streamed, and can be deployed on local GPUs, self-hosted B200 servers, Docker, or serverless Modal, with an online demo also provided.
-
Specific models and capabilities involved in project updates FastH3 V2 can run on a single consumer-grade machine, covering NVIDIA RTX 5090, RTX 4090, RTX PRO 6000 GPUs, DGX Spark, and Apple Silicon; released at the same time, FastH3 Trim is an experimental pruned model that is 4.2 times smaller than the base H3 and can run with as little as 8 GB of GPU memory. FastVideo Supports Kandinsky Lab's Kandinsky 6, enabling text-to-video and image-to-video with synchronized audio (including the base version and the 10-step distilled pi-Flow checkpoint), and supports up to 4x video super-resolution. FastH3 8-Step V2 is an eight-forward, data-free DMD2 checkpoint distilled from MiniMax-H3, featuring 80% Video Sparse Attention. FastH3 can already run locally on Apple Silicon via MLX, and on NVIDIA DGX Spark via CUDA 13, including dual-Spark inference. FastH3 Preview v1 is an open-weight 4-step sparse distilled MiniMax-H3 model for synchronized video and audio generation, developed by the project team in collaboration with Nuva Lab and the NVIDIA FastGen team. FastVideo supports MLX on Apple Silicon via FastMetal-QAD; this series includes 1.3B, 5B, and 14B models, optimized specifically for Mac. FastWan-QAD can achieve end-to-end generation of a 5-second video in 1.8 seconds. Previously, the CausalWan2.2 I2V A14B Preview model, the FastWan model and the fine-tuning and inference code for Sparse-Distillation and VSA, FastVideo V1, and the inference code for Sliding Tile Attention were also released.
III. Installation and Usage The project recommends using uv to create a clean environment, and suggests switching from Conda to uv for faster and more stable installation. On CUDA 13, UV_TORCH_BACKEND=cu130 must be used. Apple Silicon users should refer to the MLX installation guide, and can run uv pip install -e '.[mlx]' from the cloned repository, then choose a recipe in the Cookbook. On NVIDIA DGX Spark (GB10 / ARM64 + CUDA 13), since there are no prebuilt ARM wheels, installation from source in editable mode is required. The project provides installation prompts for AI coding agents; paste the prompt tolet the agent detect the platform and follow the corresponding guide. For usage, the project provides a minimal example: after creating example.py, run the script to generate a video; note that num_gpus=1 will run the worker in-process (weights are loaded only once, with no extra Python process). On machines with about 16GB of host memory such as Colab/Kaggle, num_gpus=1 should be kept.
IV. Applicable Scenarios and Ecosystem This project is suitable for developers and researchers who need video generation acceleration, post-training/distillation research, and real-time video generation and editing. The project is used by or serves as a foundation for multiple works, including SGLang's diffusion inference feature (a fork based on FastVideo), DanceGRPO, SRPO, DCM, HY-WorldPlay, as well as the SSTA proposed by Hunyuan Video 1.5 based on Sliding Tile Attention, Kandinsky-5.0's NABLA attention including a Sliding Tile Attention branch, and LongCat Video's block sparse attention, which is similar to Video Sparse Attention. The project welcomes contributions and provides a development roadmap.
V. Limitations and Notes The current materials do not provide specific pricing models, prices, or commercial licensing information; they do not provide a complete list of supported models, so the project's support matrix page should be consulted; hardware compatibility and optimization compatibility during installation and usage should be based on the project's dedicated pages; although Dreamverse's deployment method is listed, the specific configuration details are not elaborated in the materials.
