Open-H3-Regenerate-2K upscales MiniMax-H3 drafts to 2560×1440
Original title: Efficient-Large-Model/Open-H3-Regenerate-2K — new model trending #30 on Hugging Face
Authors report a five-second clip falling from 87.26 seconds to 22.45 seconds on one B200 with the fast decoder.
Efficient-Large-Model published Open-H3-Regenerate-2K, an open stand-in for a closed API that regenerates a 768P MiniMax-H3 draft at 2560×1440. The bundle merges three LoRAs into one rank-576 adapter and includes a fast non-causal H3 VAE decoder. On one B200, the authors' warm median fell from 87.26 seconds with dense attention to 22.45 seconds using Sol-Attn, MXFP8 and that decoder. The refiner and base weights follow the MiniMax H3 Community License, while other included components keep their own licenses.
Unconfirmed: Speed figures are the authors' median of three warm runs, not an independent benchmark.
- The draft is 1344×768 and 121 frames, refined in three Euler steps at CFG 1.
- The merged LoRA matches the three adapters to a relative error of 2.2e-8.
- The fast decoder has 1.62 billion parameters and 24 non-causal layers.
- Base weights are an unmodified MiniMax-H3 copy of revision 6818f6c.
- Vendors
- MiniMaxEfficient-Large-Model
- AI models
- Open-H3-Regenerate-2KMiniMax-H3
From the original article
2 min read at huggingface.coAn open replacement for the closed H3-Regenerate-2K API: a three-step MiniMax-H3 refinement pass that regenerates a [Project page](https://nvlabs. github. io/Sana/Open-H3-Regenerate-2K/) · [Code](https://github.
By Efficient-Large-Model · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.