NVIDIA and MIT team releases Open-H3-Regenerate-2K, regenerating MiniMax H3 video drafts at up to native 4K

The Efficient-Large-Model team, made up of NVIDIA and Massachusetts Institute of Technology researchers, has released the weights for Open-H3-Regenerate-2K on Hugging Face. The project is built on MiniMax H3 and regenerates low-resolution video drafts as high-resolution video. It outputs 2K (2560×1440) and supports native 4K (3840×2160) regeneration at the upper end. Inference code is available in the NVlabs/Sana repository.

The details come from a Chinese-language article on QbitAI (量子位), which republished the piece with authorization. The article was provided by MiniMax.

Where it fits in the H3 stack

According to the article, the full MiniMax H3 system covers multimodal context processing, base audio-video generation and high-resolution regeneration. The open-weight H3-Base generates 768P audio-video. The official H3-Regenerate-2K module is served through an API. Open-H3-Regenerate-2K is a self-deployable open implementation of that high-resolution regeneration stage, reusing H3's base generation capability.

How it works

During training, the team kept the H3 Transformer backbone and the video VAE parameters unchanged and trained only a new refinement LoRA. At inference, that LoRA is merged with community adapters, LightX2V FL2VA Turbo and LMS Sharp. It regenerates detail at the target resolution in three denoising steps. The draft's scene, layout and motion information serve as the basis for the process.

To cut cost, the project uses a community-developed latent upscaler that enlarges the video's latent representation directly. This avoids repeated decoding and encoding at high resolution. The team also applies SoL-Engine sparse attention, low-precision computation and a fast VAE decoder to the refinement and decoding stages.

Reported performance

The project team's tests ran on a single NVIDIA B200 GPU after warmup, on a video of about 5 seconds. Latent upscaling, three-step refinement and 2K decoding together took 22.45 seconds, down from 87.26 seconds for an unoptimized implementation. The team puts this at about 3.9 times the original speed. The fast decoder cut decoding time from about 15 seconds to about 1.6 seconds.

On quality, the team compared the two systems on RefinerBench-150, using the same inputs and the same 2K output:

• UniPercept (perceptual quality, average): Open-H3-Regenerate-2K scored 60.0, and the official H3-Regenerate-2K API scored 53.4.
• VBench: The open project scored higher on aesthetic quality, imaging quality and subject consistency. The official API remained slightly ahead on motion smoothness and dynamic degree.

H3 ecosystem context

MiniMax opened H3 weights in August this year. According to the article, H3 passed 24 million downloads within a month and produced more than 300 public derivative models. The article says this made H3 one of the most downloaded models globally in 2026.

The same NVIDIA and MIT team has also released StreamRender-H3, a streaming video rendering model built on H3 Ref2VA, and Sol-H3-VAE, a fast decoder distilled from the H3 video VAE. Other H3 derivatives named in the article:

• LightX2V's MiniMax-H3 Turbo, with four-step and eight-step distilled versions
• The community fine-tune Singularity, which targets motion, facial detail and visual style
• FastVideo's FastH3 four-step preview
• Pika H3 Flex, which uses six-step sampling
• ComfyUI and Unsloth adaptations and quantized versions for local workflows

What this means

The high-resolution regeneration stage of H3 was previously reachable through MiniMax's official API. Open-H3-Regenerate-2K gives developers a version they can deploy themselves, with weights and inference code published. Because the base model and VAE stay frozen and only a LoRA is trained, the added component is the refinement layer rather than a new base model.

The speed and quality figures come from the project team's own tests. The speed figure applies to a B200 after warmup on a roughly 5-second clip. The quality comparison uses RefinerBench-150, and the official API still leads on motion smoothness and dynamic degree. Teams weighing the open version against the API should check those motion metrics against their own content.