Introduction

MiniMax H3 makes video in two moves: the open base model drafts a clip at 768P, then a hosted H3-Regenerate-2K API redraws it at 2K so faces, fur, and fabric hold up full-size. Only the first move was open. Now NVIDIA's Sana team (NVlabs) has published Open-H3-Regenerate-2K on Hugging Face, an open refiner for that second step. It takes a low-resolution video draft and regenerates it at 2560×1440 in just three diffusion steps, and it can go all the way to native 4K.

The weights went up on October 26 under the Efficient-Large-Model org, with code in the sol-engine branch of the Sana repo and a full project page with side-by-side comparisons. It runs today as a Python CLI on your own (very large) GPU, under the MiniMax H3 Community License.

What shipped

The bundle keeps the H3 transformer and video VAE frozen and adds one merged LoRA plus a few helpers:

  • Refiner LoRA: one merged file (rank 576 across 312 linear layers) that combines NVIDIA's Open-H3 refiner with two community adapters, LightX2V's FL2VA Turbo v1.2 and the LMS Sharp LoRA. The merge matches the three separate adapters to a relative error of 2.2e-8.
  • Latent upscaler: a community 3D-convolution network that enlarges the draft's latent straight onto the target grid, so there's no decode and re-encode round trip.
  • Fast H3 VAE decoder: a 24-layer, 1.62B-parameter decoder distilled from the released H3 VAE (borrowed from NVIDIA's StreamRender-H3), which decodes a 2K clip in about 1.6 seconds instead of 15.
  • Base model copy: an unmodified copy of the MiniMax-H3 Diffusers components, or you can pull them from MiniMax's official repo instead.

How it works

The draft is encoded once by the H3 VAE, the latent upscaler stretches it to the output size, then the latent is partly re-noised and cleaned up in three Euler steps with guidance off. Upscaling carries over layout, color, and motion; the three diffusion steps are what actually draw new detail at full size.

The flexibility is the fun part. The official API only accepts unmodified H3 768P output and only returns 2K. The open version takes drafts at 384P, 512P, 576P, or 768P, from H3 or from other video models (the project page shows LTX-2 and SANA-Video drafts), and outputs native 768P, 1080P, 2K, or 4K. Clips from 5 to 15 seconds refine in a single pass.

Speed and quality

NVIDIA's numbers are for one 5-second clip going from a 768P latent to 2560×1440 on a single B200:

  • Standard Diffusers path with dense BF16 attention: about 87 seconds
  • SoL-Engine path (sparse Sol-Attn, MXFP8, CUDA graphs) with the released VAE: about 37 seconds
  • SoL-Engine path with the fast H3 VAE decoder: about 22.5 seconds, roughly 3.9× faster
  • Two refine steps instead of three: upscale plus refine drops to about 14 seconds

On their RefinerBench-150 test, with both outputs at 2560×1440, the team reports that the open refiner beats the official API on all three UniPercept axes and on VBench aesthetic quality, imaging quality, and subject consistency. In a blinded user study, it was preferred in 87% of 200 judgments.

One honest caveat: this is datacenter-class today. The fast SoL-Engine path needs about 150 GB on one GPU (B200/GB200), and the pipeline loads a 32B Qwen3-VL text encoder before the video model. There's no ComfyUI node in the release yet, so for most of us the near-term win is community ports and hosted runs built on these open weights.

Try it

Setup is Linux with Python 3.11 and PyTorch 2.11 on CUDA 13.0. From the models/Open-H3-Regenerate-2K folder of the Sana repo's sol-engine branch:

pip install -r requirements.txt
python download.py --weights ./weights --hyperflow
python infer.py --weights ./weights --input-video draft.mp4 --prompt "A red fox trots through fresh snow" --output fox-2k.mp4

Add --resolution 4k for native 3840×2160, --steps 2 for the faster two-step pass, or --fwd standard if you don't have the sparse-attention backend installed. For drafts from other video models, the team recommends --input-size 576p.

Why it matters

Open video models have been great at drafting and weaker at the final polish, because the polish step usually lives behind someone's API. Open-H3-Regenerate-2K moves that step into the open: you can run it, fine-tune around it, and point it at drafts from any model you like. Expect ComfyUI builders to start wrapping it quickly, since the latent upscaler and LMS Sharp LoRA inside it already come from the Comfy community.

Original Source

Conclusion

H3's 2K finish just stopped being a black box. If you've got the hardware, grab the weights and push your favorite 768P drafts to 2K or 4K. If you don't, bookmark the repo, because this is the piece the community will shrink down next.

Want to make your own H3 drafts first? Try this on Gen → https://artrealmai.com/gen?utm_source=magazine&utm_campaign=gen&utm_content=open-h3-regenerate-2k-nvidia-open-2k-4k-video-refiner

—Aurelia ♡