Open-Sora Plan
Peking University's open-source push for high-resolution, long-duration video generation, now at v1.5.0 with an 8B model rivaling HunyuanVideo.
About
Open-Sora Plan is the PKU-YuanGroup community project trying to do openly what Sora did behind closed doors: high-resolution, long-duration video from text, images, and structural controls. Its stack is distinctive, built around a Wavelet-Flow VAE for efficient compression and a Joint Image-Video Skiparse Denoiser, plus prompt refinement, sparse attention, and bucket training to keep memory in check (93 frames at 480p inside 24GB of VRAM as of v1.3). The project moved fast: v1.2.0 added image-to-video in August 2024, v1.3 shipped the paper in December 2024, and v1.5.0 (June 2025) introduced the sparse SUV architecture, claiming HunyuanVideo-comparable quality from an 8B-scale model trained on 40 million videos.
With 12.2k GitHub stars and MIT-licensed code, it is one of the most watched open video efforts, and the team has since spun out Helios for minute-scale generation at 19.5fps on a single H100. The 'comparable to HunyuanVideo' claim is author-reported with no independent benchmark, and the v1.5.0 release was initially Ascend-NPU-first with the GPU version trailing, so CUDA users should check branch status before committing.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Author-reported (v1.5.0): performance comparable to HunyuanVideo (open-source) with an 8B-scale model trained on 40M video samples; no independent published benchmarks
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (metrics tier, metrics-v1). Evidence: PKU-YuanGroup paper arXiv 2412.00131 (Dec 2024); GitHub pku-yuangroup/open-sora-plan (12.2k stars, MIT confirmed via GitHub API, updated Oct 2026); versioned releases v1.2.0 through v1.5.0 with reports. Strengths: large community, permissive MIT license, transparent version history, distinctive efficient architecture. Docked for: (1) headline quality claim is author-reported only; (2) latest release was Ascend-first, GPU support lagging. License: MIT, confirmed via GitHub API.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 9.0/10
- Maintenance · recency, releases, issue load
- 8.0/10
- Openness · license + docs
- 9.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · PKU-YuanGroup; paper arXiv 2412.00131, Dec 2024; 12.2k GitHub stars
- · v1.2.0 Aug 2024, v1.3 Dec 2024, v1.5.0 Jun 2025 (8B SUV, Ascend-first)
- · Wavelet-Flow VAE + Skiparse DiT; 93 frames at 480p in 24GB VRAM
- · HunyuVideo-comparable claim is author-reported only