Open-H3-Regenerate-2K-IC: H3 Drafts to 2K That Keep Faces and Text

Open-H3-Regenerate-2K-IC is the second version of NVlabs' open MiniMax H3 refiner. It does the same job as the October release: lift a 1344x768 draft to 2K or 4K in three steps. The change is where the draft goes. The first version re-noised the draft and polished it, so faces, mouth shapes and small lettering drifted. The IC version keeps the draft clean in the Ref2VA transformer's context and draws the high-resolution clip from noise beside it, so identity, lip sync, sign text and layout follow the draft.

Frames from the project's own comparison clip. Source: project page

The IC output for that clip at 2560x1440, with the draft's audio. Source: project page

What it keeps

Against the official H3-Regenerate-2K API, the project's car-park comparison is the clearest case: the API renders the sign as unreadable marks, the IC version still reads "PRIVATE CAR PARK". Against the first open version, the IC output adds less new texture but leaves eyes, wrinkles and mouth shapes where the draft put them.

The IC LoRA is trained on the frozen Ref2VA transformer, which already reads reference videos as extra tokens. Each training pair is a real 768P-to-4K clip; its resized-down copy sits in front as the clean reference while the target is noised anywhere from pure noise to clean, and the loss lands on the target tokens only.

How it runs

Give it the 768P draft as <Video 1> plus the caption that generated it, unchanged. Three samplers, all at CFG 1:

• up3 (default): the draft latent goes through the 3D-conv latent upscaler, is noised to sigma 0.835 and finished in three transformer calls.
• up4: same upscaled start at sigma 0.923, four calls.
• hf8: pure noise and all eight HyperFlow steps.

Output is 1080p, 2K (2560x1440, default) or 4K (3840x2160); 384P, 512P and 576P drafts come back at 768P. Audio is copied from the draft unchanged, a silent audio latent ships with the weights, and the 2K result decodes through the fast H3 VAE while 4K uses the standard one.

In ComfyUI

There are no official nodes, but two community routes exist already:

• Kijai compressed the 8.3 GB FP32 LoRA to a 200 MB rank-21 bf16 file that loads as a normal LoRA on the Ref2VA model.
• ComfyUI-OpenH3Regenerate2KIC ships four nodes (Prepare Draft, Official Prompt, Sigma Grid, Compose Target Latent), the full-rank 5.9 GB converted LoRA and three workflows: text to draft to 2K, reference image to draft to 2K, and draft to 2K.

Weights and inference code are on Hugging Face, with the reference CLI in NVlabs/Sana.