Open-H3-Regenerate Unlocks 4K Output From MiniMax-H3's API-Locked Stage
An open in-context LoRA on Hugging Face reconstructs MiniMax-H3 drafts at 2K or 4K in three to four steps, filling a gap MiniMax left closed.
- Open LoRA that regenerates MiniMax-H3 drafts at 1080p, 2K or 4K
- Uses in-context conditioning on the draft via H3's Ref2VA transformer
- Preserves faces, lip sync, on-screen text and scene layout
- Only 3 or 4 denoising steps with CFG 1, via HyperFlow distillation
- Ships IC LoRA, latent upscaler, fast VAE and inference code
- Covered by the MiniMax H3 Community License with territorial restrictions
MiniMax’s open H3 checkpoints stop at 768p because the company’s H3-Regenerate-2K stage remains available only through its API. Open-H3-Regenerate-2K-IC, published on Hugging Face by Efficient-Large-Model, supplies an open alternative for that missing stage. It conditions H3 on a low-resolution draft and regenerates the clip at 1080p, 2K, or 4K in three or four denoising steps, according to the release documentation.
H3’s missing final pass
MiniMax published H3 under its community license as a 33.1-billion-parameter video model. It can generate clips up to 15 seconds long with native 32 kHz stereo audio. A Qwen3-VL-32B text encoder interprets the prompt, while a single-stream transformer jointly models the video and audio output.
The released H3 checkpoints generate at up to 768p. Open-H3-Regenerate-2K-IC adds the high-resolution pass through a LoRA adapter for H3’s reference transformer, a latent upscaler, a faster VAE decoder, and an inference pipeline. It still requires the MiniMax-H3 base weights.
A draft becomes the context
The initials “IC” refer to in-context generation. During inference, the low-resolution clip enters H3’s transformer_ref component as a reference video. A LoRA, which is a set of task-specific weight updates layered onto the frozen model, then guides the transformer to generate a higher-resolution version of the same scene.
Keeping the draft inside H3’s generation context gives the model access to faces, mouth movements, text, composition, and timing while it reconstructs detail. The pipeline accepts the original generation prompt unchanged, and the caption can also be left empty.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
