Open-Sora

Open-Sora is an open-source high-quality video generation project from HPC-AI Tech (LuChen). Its mission is to bring Sora-class video generation into the community with full open-sourced training code and weights. It follows a DiT (Diffusion Transformer) approach — 3D-VAE plus rectified flow plus score conditioning — and supports text-to-video, text-to-image, image-to-video, video-to-video, and infinite-length extension.

The core line has gone through four iterations. 1.0 (2024-03-18) shipped 2-second 512x512 video generation after 3 days of training. 1.1 (2024-04-25) added 2-15 second clips at 144p to 720p with arbitrary aspect ratios and exposed the full processing pipeline. 1.2 (2024-06-17) introduced 3D-VAE and rectified flow, lifting video quality noticeably. 1.3 (2025-02-20) upgraded both the VAE and the Transformer architecture. 2.0 (2025-03-12) shipped the 11B model — the VBench gap to OpenAI Sora narrowed from 4.52% to 0.69%, and human-preference blind tests tied HunyuanVideo 11B and Step-Video 30B, with the full training cost landing at only $200,000.

Use cases include research teams reproducing Sora-style training pipelines, enterprises self-hosting video generation services, and independent creators running 5-15 second text or image-to-video clips locally. Deployment relies on Python, PyTorch, and CUDA; H100 or H800 GPUs are recommended. At 768x768, a single GPU finishes 50 inference steps in 1656 seconds, dropping to 276 seconds on an 8-GPU box; at 256x256 a single GPU produces a clip in roughly 60 seconds. Model weights are mirrored on Hugging Face and ModelScope at hpcai-tech/Open-Sora-v2, and a Gradio demo is included for quick experimentation. Licensed under Apache-2.0; currently 29k+ stars and 3k+ forks on GitHub. Worth noting: the 2.0 line now feeds the commercial Video Ocean product and the hpc-ai.com model API, so open-source repository activity has slowed — the most recent commit is 2026-04-09.