Introduction

NVIDIA's SoL-Engine team (the NVlabs Sana group) just shipped Open-H3-Regenerate-2K-IC, an open-weight video-to-video refiner for MiniMax H3. It takes a low-res H3 draft and regenerates it at 1080P, 2K (2560×1440) or 4K (3840×2160) while keeping faces, lip sync, lettering and layout locked to the draft. Weights landed today on Hugging Face, with inference code in the NVlabs/Sana repo and a write-up on the project page. You run it locally on your own GPUs; it's an IC LoRA on top of H3, not a hosted API.

What's New vs. the October Release

We covered the first Open-H3-Regenerate-2K on October 28: a three-step refiner LoRA that re-noises the draft and redraws it at full size. It added lovely texture, but the team admits it could drift on eyes, wrinkles, mouth shapes and text.

The IC ("in-context") version fixes that by changing where the draft lives:

  • The draft is kept clean in the transformer's context as a reference video, instead of being noised and repainted
  • It uses H3's own Ref2VA reference-to-video transformer (transformer_ref), frozen, with a new IC LoRA (rank 384, 312 linear layers) trained on top
  • The high-res clip is generated next to the draft, and every step attends to it, so faces, lip sync, signage and layout follow the original
  • The draft's audio track is copied over unchanged

On the project page's side-by-sides, the official H3-Regenerate-2K API turns a "PRIVATE CAR PARK" sign into unreadable marks, while the IC version keeps it legible.

Key Specs

  • Input: a MiniMax H3 draft at 768P (1344×768), read at 24 fps, 124 frames
  • Output: 1080P, 2K (default) or 4K; also 384P/512P/576P drafts up to 768P via --lq
  • Samplers: up3 (3 steps, default), up4 (4 steps), or hf8 (8 HyperFlow steps from pure noise), all at CFG 1
  • Bundled adapters: HyperFlow 8-step LoRA and LMS Sharp v1.0 merge in at load time; the IC LoRA sits on top
  • Decoder: fast H3 VAE decoder v2 (24 layers) by default, or the standard H3 VAE
  • Multi-GPU: sharded with FSDP plus context parallelism over 1, 2, 4, 7 or 8 GPUs
  • Prompt: pass the exact H3 prompt that made the draft, or leave it empty

How to Run It

Install and download

The code targets Linux, Python 3.11, CUDA PyTorch, Diffusers from main, and ffmpeg on your path.

git clone -b sol-engine https://github.com/NVlabs/Sana.git
cd Sana/models/Open-H3-Regenerate-2K/IC
pip install -r requirements.txt
python download.py --out ./weights

Regenerate a draft at 2K

torchrun --nproc_per_node 8 infer.py --weights ./weights \
  --input draft.mp4 --output draft_2k.mp4 \
  --caption "the MiniMax-H3 prompt the draft was generated from"

Swap in --resolution 4k for 4K, --sampler hf8 for the full eight-step path, or --jobs jobs.jsonl to batch several clips while loading the models once. Set --nproc_per_node to match your GPU count (1, 2, 4, 7 or 8).

Using it in other runtimes

If your tool only loads one LoRA per model, python merge_lora.py --weights ./weights folds the IC LoRA and LMS Sharp into a single exact LoRA; load HyperFlow separately before it. There's no official ComfyUI node in this drop yet, so this is a script-first release for now.

License Check

The release ships under the MiniMax H3 Community License, which excludes the European Union, the United Kingdom, South Korea and the United States. Bundled pieces keep their own terms (the latent upscaler is Apache-2.0 weights with MIT code, and the Qwen3-VL text encoder and LMS Sharp LoRA carry their own licenses). It's also not an official MiniMax product, so read the LICENSE and NOTICE files before shipping commercial work.

Original Source

Conclusion

The first open regenerate pass gave H3 owners a way to reach 2K and 4K on their own hardware; this IC version makes that upgrade trustworthy for talking heads, product shots and anything with text on it. Draft fast at 768P, keep the take you love, then let the IC LoRA paint in the detail without changing the performance.

Try this on Gen → https://artrealmai.com/gen?utm_source=magazine&utm_campaign=gen&utm_content=open-h3-regenerate-2k-ic-nvidia-consistent-4k-refiner

—Aurelia ♡