R_REDDYX.XYZ
VideoPython★ 4,831⑂ 528 forksApache-2.0

hao-ai-lab/FastVideo

UNIFIED FRAMEWORK FOR ACCELERATED VIDEO GENERATION

FastVideo is a unified framework for post-training and inference of accelerated video generation. It supports distillation, sparse attention, and real-time via Dreamverse.

// KEY FEATURES

  • Post-training: full fine-tuning and LoRA for video DiT
  • Sparse distillation DMD2: >50× denoising speedup
  • Inference with Sequence Parallelism and multiple attention backends
  • Dreamverse: real-time video generation and editing

// DEEP DIVE: HOW IT WORKS

FastVideo combines a post-training pipeline (full fine-tuning, LoRA, DMD2 distillation, Self-Forcing) with optimized inference using Sequence Parallelism, multiple attention backends, and CLI/API. It supports bidirectional and autoregressive video DiTs, includes preprocessing for video/images/text, and scalable training via FSDP2 with activation checkpointing.

It stands out with acceleration techniques: Video Sparse Attention, sparse distillation achieving >50× denoising speedup, and Dreamverse for real-time streaming and "vibe directing" of video. The repo integrates PAN 2, FastH3 Omni, Kandinsky 6, FastH3 V2/Trim models with base and distilled checkpoints.

Requires Python 3.12, CUDA 12 (cu126) for NVIDIA GPUs (H100, A100, 4090/5090, RTX PRO 6000, DGX Spark), runs on Linux, Windows, macOS (Apple Silicon via MLX) and AMD/ROCm. Deployment via Docker, Modal, SSH to B200. Minimum VRAM for FastH3 Trim is 8 GB.

// USE CASES

Best for: For ML engineers and researchers to accelerate video generation on GPU

  • Train accelerated video generation model via sparse DMD2 distillation
  • Run video model inference with Sequence Parallelism across multiple GPUs
  • Deploy Dreamverse for real-time video streaming and editing

// QUICK START

uv venv --python 3.12 --seed
source .venv/bin/activate
UV_TORCH_BACKEND=cu126 uv pip install fastvideo

Commands are taken from the repository README — check the current version on GitHub before running.

#diffusers#diffusion-models#distillation#inference#post-training#video-generation
Open on GitHub →

New repositories every 30 minutes

REDDYX AI scans GitHub 24/7 and ships the best AI/ML/Web3 projects to Telegram.

Join on Telegram
← Full catalog·Full index

// SIMILAR REPOSITORIES

6700+ PROMPTS FOR SEEDANCE 2.0★ 223651 VIDEO EFFECTS FOR CLAUDE CODE★ 969OPEN-SOURCE IPTV PLAYER FOR ANDROID TV★ 790OPEN-SOURCE AI VIDEO EDITOR FOR CLAUDE CODE★ 609VIDEO ENGINE AS CODE WITH AGENT SKILLS★ 209AUTOMATIC VIDEO TRANSLATION WITH DUBBING★ 19764

← FULL CATALOG

hao-ai-lab/FastVideo — Unified Framework For Accelerated…