NVIDIA just released Open-H3-Regenerate-2K-IC, an open-weight AI video refinement model that upgrades low-resolution MiniMax H3 videos to 2K and 4K while preserving faces, lip sync, text, and scene composition.
And it can generate the high-resolution result in just THREE diffusion steps.
This is NVIDIA Research's alternative to MiniMax's proprietary H3-Regenerate-2K API, with a fundamentally different approach to video consistency.
Traditional AI video refinement often adds impressive details but can subtly change faces, mouth movements, eyes, and even text between frames.
Open-H3-IC addresses this problem through In-Context generation.
Instead of repeatedly adding noise to the original video, it keeps a clean version of the source inside MiniMax H3's Ref2VA transformer.
The high-resolution video is then generated alongside that reference, allowing the model to continuously attend to the original footage.
Here's what's new:
• In-Context video refinement: the original video remains accessible to the transformer throughout generation.
• Better facial consistency: aims to preserve facial identity, expressions, eye details, and lip movements.
• Improved text preservation: reduces the risk of signs and lettering turning into unreadable artifacts.
• Native 4K output: supports 1080p, 2K (2560×1440), and 4K (3840×2160) regeneration.
• Three-step diffusion: the default up3 sampler uses only three transformer passes to refine the video.
• Latent-space upscaling: enlarges video representations before generating new high-resolution detail.
• Ref2VA architecture: combines a clean low-resolution reference with high-resolution target generation.
• Audio preservation: keeps the original video's soundtrack unchanged.
• Open implementation: NVIDIA has published inference code and model weights for local deployment.
• Flexible sampling: choose three, four, or eight diffusion steps depending on the workflow.
The real breakthrough isn't simply making AI videos sharper.
It's generating new detail while keeping the original video as a persistent reference.
For AI filmmakers and VFX artists, that could mean fewer unwanted facial changes, more reliable lip sync, and better consistency during high-resolution refinement.
But this isn't magic.
Face drift and other artifacts are reduced, not mathematically eliminated. The workflow also requires substantial GPU resources, and the model weights are subject to the MiniMax H3 Community License.
The frequently cited 5x speed comparison refers to NVIDIA's earlier Open-H3-Regenerate-2K model, not a separately validated benchmark for this new IC version.
NVIDIA Research project and comparisons:
https://nvlabs.github.io/Sana/Open-H3-Regenerate-2K/#ic
GitHub source code:
https://github.com/NVlabs/Sana/tree/sol-engine/models/Open-H3-Regenerate-2K/IC
Model weights:
https://huggingface.co/Efficient-Large-Model/Open-H3-Regenerate-2K-IC
SoL-Engine:
https://nvlabs.github.io/Sana/Sol-Engine/
Would you use an open AI video upscaler instead of a proprietary cloud API?
#NVIDIA #AIVideoUpscaling #MiniMaxH3
YouTube / ABVX
