Nvidia and MIT's Open-H3-Regenerate-2K lifts MiniMax H3 videos to 2K and 4K

What this is about

The "Efficient-Large-Model" team, made up of researchers from Nvidia and the Massachusetts Institute of Technology (MIT), has released the model weights for the video project Open-H3-Regenerate-2K on Hugging Face. The matching inference code is in the GitHub repository NVlabs/Sana. The project builds on MiniMax H3, the video generation model from the Chinese company MiniMax. It turns low-resolution video drafts into high-resolution video: 2K (2560×1440) and up to native 4K (3840×2160). The report comes from QbitAI and, according to its text, was provided by MiniMax. The numbers are therefore claims from the project team or from MiniMax's circle.

How MiniMax H3 is built

According to the main source, the overall H3 system has several stages: processing multimodal inputs (text, image, video, audio), base generation of audio and video, and high-resolution regeneration. The base model H3-Base has open weights and generates audio and video at 768p. The official H3-Regenerate-2K module, by contrast, is only available through an API. The new project replaces exactly this last stage with an open variant you can run yourself.

Some background (per intelligentliving.co): H3 is an "omni-modal" model with a 33-billion-parameter transformer (parameters are, roughly speaking, the knobs a model learns during training). The full system, including the Qwen3-VL-32B text encoder, comes to about 69.2 billion parameters. It produces clips of 4 to 15 seconds at 24 frames per second with synchronized stereo sound. Instead of a classic super-resolution module, which often smears detail, H3 regenerates its own low-res output "in context." That is the approach the Nvidia/MIT team reproduces.

How the refinement works

When training the refinement model, H3's main transformer and the video VAE (an autoencoder that packs video into a compact representation, the "latent space," and converts it back) stayed unchanged. Only one additional LoRA was trained. That is a small add-on weight attached afterward, which adapts a large model with little effort.

At inference, this refinement LoRA is merged with community adapters such as LightX2V FL2VA Turbo and LMS Sharp. In just three denoising steps (the steps in which an image gradually emerges from noise), the details are regenerated at the target resolution. The scene, layout and motion of the draft serve as the foundation.

Speed

To go faster, the project uses a community-built latent upscaler that enlarges the latent representation directly. That avoids repeated decoding and encoding at high resolution. Sparse attention (the attention mechanism computes only on selected parts), low-precision computing and a fast VAE decoder from the SoL-Engine software also speed up refinement and decoding.

That works out to roughly 3.9 times the speed. Note that the B200 is a data-center GPU. These timings don't apply to home hardware.

Quality compared with the official API

On the RefinerBench-150 test set, the team compared both solutions with the same input and the same 2K output. On the UniPercept perceptual metric, Open-H3-Regenerate-2K averages 60.0, while the official H3-Regenerate-2K API scores 53.4. In VBench, the open approach leads on aesthetic quality, image quality and subject consistency. The official API remains slightly ahead on motion smoothness and dynamics. These are the project team's own measurements, and the sources include no independent verification.

The H3 ecosystem

According to the main source, MiniMax H3 was released with open weights in August. It was reportedly downloaded more than 24 million times within a month, has over 300 public derivatives, and ranks among the most downloaded models of 2026. The same Nvidia/MIT group also released StreamRender-H3 (streaming video rendering based on H3 Ref2VA) and Sol-H3-VAE (a fast decoder distilled from the H3 VAE). Other community projects include LightX2V with MiniMax-H3 Turbo (4-step and 8-step variants), the "Singularity" fine-tune, FastH3 from FastVideo, Pika H3 Flex with six steps, and ComfyUI and Unsloth adaptations for local workflows.

License: a warning for Europe

According to intelligentliving.co, H3 is released under the "MiniMax H3 Community License." It excludes the US, the EU, the UK and South Korea from the "Applicable Territory." People there reportedly may not use the weights locally or use their outputs. The hosted API remains available worldwide. The sources don't say how the license applies to the derived Open-H3-Regenerate-2K. If you work in Germany or elsewhere in the EU, check this before you use it.

What you need to try it

The sources give no minimum requirements for a home system. Testing ran on an Nvidia B200 (data center). For 2K/4K regeneration with a 33-billion-parameter backbone, a large GPU with plenty of VRAM (video memory) is realistic. One community video does show a ComfyUI workflow for H3 refinement. Running it on a CPU or a Mac makes little practical sense. Plan for tens to well over a hundred gigabytes of disk space for the H3 base, VAE and adapters (an estimate, not from the source).

Takeaway

The project fills a gap. H3's high-quality resolution stage was previously available only via API. Now there is an open, customizable alternative that, by its own measurements, scores better on perceived image quality. Nvidia's involvement shows how firmly H3 is establishing itself as a platform for community development. On the critical side, all the numbers come from the project team or from a report supplied by MiniMax, independent tests are missing, and the license question for users in the EU is open at best and unfavorable at worst. The speed figures apply only to top-end hardware.