Recently, the Efficient-Large-Model team composed of researchers from NVIDIA and MIT released the model weights of the video enhancement project Open-H3-Regenerate-2K on Hugging Face, and the accompanying inference code is now available in the NVlabs/Sana repository. The solution is based on MiniMax H3 and regenerates low-resolution video drafts into high-resolution videos, supporting 2K output at 2560×1440 and achieving native 4K regeneration of up to 3840×2160.
The complete MiniMax H3 system includes multimodal context processing, base audio-video generation, and high-resolution regeneration stages. Among these, H3-Base, whose weights have been released, generates 768P audio-video, while the official H3-Regenerate-2K module is provided as a service via API. The project released this time reuses the generation capability of the H3 base model, providing an open, self-deployable implementation for the high-resolution regeneration stage.
During the training of the refiner model, the research team kept the original parameters of the H3 Transformer backbone and the video VAE unchanged, training only the newly added refiner LoRA. At inference time, the refiner LoRA is merged with community adapters such as LightX2V FL2VA Turbo and LMS Sharp, regenerating picture details at the target resolution through three-step denoising, with the scenes, layouts, and motion information in the video draft providing the foundation for this process.
To improve inference efficiency, the project uses a community-developed latent-space upsampler that directly upscales the video's latent representations, reducing the overhead of repeated decoding and encoding at high resolutions. Meanwhile, the solution combines SoL-Engine's sparse attention, low-precision computation, and fast VAE decoder to accelerate the refinement and decoding stages.
Tests published by the project team show that on a single NVIDIA B200 GPU under warmed-up test conditions, the combined time for latent-space upscaling, three-step refinement, and 2K decoding of a roughly 5-second video dropped from 87.26 seconds with the unoptimized implementation to 22.45 seconds, a processing speed of about 3.9 times the original. Among these, the fast decoder reduced video decoding time from about 15 seconds to about 1.6 seconds.
In terms of image quality, the project team conducted comparative evaluations on RefinerBench-150. Under identical input and identical 2K output conditions, Open-H3-Regenerate-2K achieved an average score of 60.0 on the perceptual quality benchmark UniPercept, while the official H3-Regenerate-2K API scored 53.4. In VBench evaluations, this solution scored higher on aesthetic quality, imaging quality, and subject consistency, while the official API remained slightly better on motion smoothness and dynamic degree.
MiniMax H3 opened its weights in August this year, and fine-tuning, distillation, and tool adaptations around H3 quickly proliferated, attracting research teams and community developers from home and abroad, with over 24 million downloads in one month and over 300 public derivative models, making H3 one of the most downloaded models worldwide in 2026.
Among these, the joint NVIDIA and MIT development around H3 covered different directions. In addition to this video enhancement solution, the team also released a streaming video rendering model StreamRender-H3 built on H3 Ref2VA, as well as a fast decoder Sol-H3-VAE distilled from the H3 video VAE, targeting interactive video rendering and decoding efficiency optimization respectively.
Earlier, LightX2V released MiniMax-H3 Turbo offering distilled versions such as four-step and eight-step variants, while the community fine-tuned model Singularity focused on optimizing motion performance, facial details, and visual style. Overseas teams FastVideo and Pika also released the four-step preview version FastH3 and the six-step sampling Pika H3 Flex respectively; ComfyUI and Unsloth brought H3 into more local creative workflows through adapted weights and quantized versions.