Open-H3-Regenerate-2K

Open-H3-Regenerate-2K is NVIDIA's open video refinement stack for MiniMax H3 drafts, published on Hugging Face under Efficient-Large-Model. It is positioned as an open replacement for MiniMax's closed H3-Regenerate-2K API: encode a low-res draft, latent-upscale, then run a 3-step H3 diffusion refine to 768P / 1080P / 2K / 4K.

HF: Efficient-Large-Model/Open-H3-Regenerate-2K · Project page: nvlabs.github.io/Sana/Open-H3-Regenerate-2K

Key points

1. Named open model: not a MiniMax version bump—NVIDIA-trained refiner LoRA + latent upsampler + optional fast H3 VAE decoder bundle.
2. Pipeline: draft → H3 VAE encode → 3D latent upscaler → re-noise → 3 Euler steps with merged refiner/Turbo/Sharp LoRA → decode (standard or StreamRender-H3 fast VAE).
3. Speed claim: ~22.45 s for a 5 s 2K clip on one B200 with Sol-Engine + fast VAE (vendor numbers).
4. Ecosystem: multiple ComfyUI custom-node ports appeared within days; CN creators comparing it to the closed API.
5. License caveat: components inherit MiniMax H3 Community License territory/use terms plus third-party LoRA licenses—not a single blanket OSS license.

Why it matters

Creators searching for an open 2K regenerate path for H3 will hit this exact coinage; distinct from already-tracked FastH3 Omni / Pika H3 Flex.

How to try it

1. Download weights from the HF repo (LoRA, sharp branches, latent upsampler, optional fast VAE).
2. Follow the NVLabs project page or a maintained ComfyUI Open-H3-Regenerate node; feed a 384P–768P draft.
3. Compare 2K / 4K outputs against the closed MiniMax API on the same draft; watch VRAM and license territory limits.

Keep reference conditioning tight—community reports note hallucinated detail when references are weak.