hao-ai-lab/FastVideo

Health v2 · maintenance only

Adoption not part of the score

4076 stars · 418 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-29, confidence not recorded

FastVideo is a unified Python framework for post-training and real-time inference of video diffusion models, covering data preprocessing, finetuning (full and LoRA), stepwise/causal distillation, quantization-aware training, and accelerated inference. It supports NVIDIA CUDA GPUs and Apple Silicon via MLX, with both CLI and Python APIs.

Use cases

• generate videos from text prompts with diffusion models
• distill a video diffusion model to run in few steps
• finetune or LoRA-train a video DiT model on custom data
• run realtime video generation on a single GPU
• quantize a video model with QAT for faster inference
• run video generation locally on Apple Silicon Macs

When to choose

• you need end-to-end video diffusion training, distillation, and inference in one framework
• you want state-of-the-art speedups like DMD2 distillation, sparse attention, or quantization
• you need distributed training/inference with FSDP2 and sequence parallelism
• you want realtime streaming video generation or editing

When to avoid

• you only need simple image generation rather than video
• you want a no-code GUI tool rather than Python/CLI workflows
• you lack a supported GPU (CUDA 12.6+ NVIDIA or Apple Silicon M-series)
• you need a lightweight inference-only library without training features

Facets

library · maturity active

machine-learning deep-learning llm-inference video-processing gpu-computing cli deep-learning machine-learning artificial-intelligence gpu-computing python cli diffusion-models video-generation distillation post-training lora-finetuning quantization sparse-attention realtime-inference diffusers apple-silicon video linux macos gpu

Member repositories

For agents

markdown · JSON · MCP: product_card(name="hao-ai-lab/FastVideo")