Film · Model Watch
NVIDIA’s H3 refiner tries to enlarge AI video without losing the face
Open-H3-Regenerate-2K-IC uses the original draft as a reference when regenerating higher-resolution video.
Reporting updated Nov 11, 2026

THE TAKEAWAY
The appeal of Open-H3-Regenerate-2K-IC is preserving continuity when a low-resolution video draft becomes a deliverable.
WHAT HAPPENED
The NVIDIA-linked research release uses in-context conditioning on an original MiniMax H3 draft while producing higher-resolution frames. Its stated goal is to retain facial identity, lip synchronization, composition and lettering more consistently than a simpler re-noise approach. Published examples and steps are first-party demonstrations; they need independent comparison. The associated MiniMax H3 licensing contains geographic restrictions that may prevent use in the US, EU, UK and South Korea without additional rights.
WHY IT MATTERS
A close-up can become unusable if the face or mouth changes after finishing. Keeping the original draft in the process is a practical answer to an expensive continuity problem.
Reader card
Draft-conditioned generation
preservation of faces and text
resolution choices
geographic license restrictions
VEYR ANGLE
VEYR must place the license caveat prominently and avoid presenting source availability as universal commercial permission.