NVIDIA and MIT teams' Open-H3-Regenerate-2K, an open-source video enhancement solution based on MiniMax H3…
The Efficient-Large-Model team, composed of researchers from NVIDIA and MIT, has released the model weights for the video enhancement project Open-H3-Regenerate-2K on Hugging Face, with the accompanying inference code now available in the NVlabs/Sana repository. Based on MiniMax H3, the solution regenerates low-resolution video drafts into high-resolution video, supports 2K output at 2560×1440, and can achieve native 4K regeneration at up to 3840×2160. The team kept the H3 Transformer backbone and video VAE parameters unchanged, training only the newly added refinement LoRA; at inference time it is merged with community adapters such as LightX2V FL2VA Turbo and LMS Sharp, regenerating image details through three-step denoising. On a single NVIDIA B200 GPU, the combined time for latent-space upscaling, three-step refinement, and 2K decoding of an approximately 5-second video dropped from 87.26 seconds to 22.45 seconds, about 3.9 times faster than before; on RefinerBench-150, UniPercept averaged 60.0, compared with 53.4 for the official H3-Regenerate-2K API.
Why read this
AI score 57
The original text provides the open weights, the three-step refinement pipeline, and measured latency on a single GPU, so readers can judge the cost and quality trade-offs of self-deploying 2K video regeneration.
Topics
Body · AI translation
Recently, the Efficient-Large-Model team, composed of researchers from NVIDIA and MIT, released the model weights for the video enhancement project Open-H3-Regenerate-2K on Hugging Face, with the accompanying inference code 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 2560×1440 2K output and achieving native 4K regeneration at up to 3840×2160.
The complete MiniMax H3 system includes multimodal context processing, base audio-video generation, and high-resolution regeneration. Among these, H3-Base, whose weights have been released, is responsible for generating 768P audio-video, while the official H3-Regenerate-2K module provides services via API. The project released this time reuses the generation capabilities of the H3 base model, providing a self-deployable open implementation for the high-resolution regeneration stage.
During the training phase of the refinement model, the research team kept the original parameters of the H3 Transformer backbone and the video VAE unchanged, training only the newly added refinement LoRA. At inference, the refinement LoRA is merged with community adapters such as LightX2V FL2VA Turbo and LMS Sharp, and through three-step denoising, regenerates image details at the target resolution. The scene, layout, and motion information in the video draft provide the basis for this process.
To improve inference efficiency, the project adopts a latent-space upscaler developed by the community to directly upscale the video's latent representation, reducing the overhead of repeated decoding and encoding at high resolution. At the same time, the solution combines SoL-Engine's sparse attention, low-precision computation, and a 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 an approximately 5-second video dropped from 87.26 seconds in the unoptimized implementation to 22.45 seconds, about 3.9 times faster than 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 the same input and same 2K output conditions, Open-H3-Regenerate-2K scored an average of 60.0 on the perceptual quality benchmark UniPercept, while the official H3-Regenerate-2K API scored 53.4. In the VBench evaluation, the 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 adaptation around H3 rapidly increased, attracting participation from research teams and community developers at home and abroad. Within one month, downloads exceeded 24 million, and more than 300 public derivative models were produced, making H3 one of the most downloaded models globally in 2026.
Among these, the joint development around H3 by NVIDIA and MIT covered different directions. In addition to this video enhancement solution, the team also released the streaming video rendering model StreamRender-H3 built on H3 Ref2VA, and the fast decoder Sol-H3-VAE distilled from the H3 video VAE, aimed at interactive video rendering and decoding efficiency optimization, respectively.
Previously, LightX2V's MiniMax-H3 Turbo provided distilled versions such as four-step and eight-step, 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 FastH3 four-step preview version 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.
This article is provided by MiniMax; QbitAI is authorized to repost it, and the views belong to the original author.