The Zhitong Finance App learned that recently, the Efficient-Efficiency Model team, composed of Nvidia and MIT researchers, released the model weights for the video enhancement project Open-H3-Regenerate-2K on Hugging Face. The supporting inference code has been opened in the NVLabs/Sana warehouse. Based on MiniMax H3, this solution regenerates low-resolution video drafts into high-resolution videos, supports 2560×1440 2K output, and can achieve native 4K regeneration of up to 3840×2160.
The MiniMax H3's complete system includes multi-modal context processing, basic audio and video generation, and high-resolution regeneration. Among them, H3-Base, which has opened weights, is responsible for generating 768P audio and video, and the official H3-Regenerate-2K module provides services through an API. The project released this time reuses the generation capabilities of the H3 basic model to provide a set of open implementations that can be deployed on its own for high-resolution regeneration processes.
During the training phase of the detailed model, the research team kept the original parameters of the H3 Transformer backbone and video VAE unchanged, and only trained the newly refined LoRa. When reasoning, refining LoRa is used in combination with community adapters such as LightX2V FL2VA Turbo and LMS Sharp to regenerate image details at the target resolution through three-step denoising. The scene, layout, and motion information in the draft video provide the foundation for this process.
In order to improve reasoning efficiency, the project uses a community-developed latent space amplifier to directly amplify the potential representation of the video and reduce the cost of repeated decoding and coding at high resolution. At the same time, the solution combines Sol-Engine's sparse attention, low precision calculation, and fast VAE decoder to accelerate refinement and decoding.
Tests published by the project team showed that under test conditions with a single NVIDIA B200 GPU and warm-up, the total time required for a 5-second video amplification, three-step refinement, and 2K decoding was reduced from 87.26 seconds without optimization to 22.45 seconds, and the processing speed was about 3.9 times that of the original. Among them, the fast decoder reduced the video decoding time from about 15 seconds to about 1.6 seconds.
In terms of image quality, the project team conducted a comparative evaluation on RefinerBench-150. Under the same input and same 2K output conditions, the average score of Open-H3-Regenerate-2K on the perception quality evaluation UniPercept was 60.0, and the official H3-Regenerate-2K API was 53.4. In the vBench evaluation, the solution scored higher for aesthetic quality, imaging quality, and subject consistency, and the official API was still slightly superior in terms of motion smoothness and dynamism.
MiniMax H3 opened its weight in August of this year, and the fine-tuning, distillation, and tool adaptation around H3 rapidly increased, attracting the participation of research teams and community developers at home and abroad. The number of downloads exceeded 24 million times within a month, generating more than 300 publicly derived models. H3 became one of the most downloaded models in the world in 2026.
Among them, the joint development of Nvidia and MIT covered different directions around H3. In addition to this video enhancement solution, the team also released StreamRender-H3, a streaming video rendering model built on H3 ref2VA, and Sol-H3-VAE, a fast decoder distilled from H3 video VAE, which are optimized for interactive video rendering and decoding efficiency, respectively.
Previously, the MiniMax-H3 Turbo launched by LightX2V provided four-step and eight-step distillation versions, while the community fine-tuning model Singularity focused on optimizing movement 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 by adapting weights and quantization versions.