NVIDIA and MIT release open-weight Open-H3-Regenerate-2K video upscaler built on MiniMax H3
Overview
NVIDIA and MIT's Efficient-Large-Model team has released the Open-H3-Regenerate-2K model weights on Hugging Face, with inference code in the NVlabs/Sana repository.
The project builds on MiniMax H3 to regenerate low-resolution video drafts at 2560×1440 (2K), with native 3840×2160 (4K) regeneration supported.
The team says its optimized pipeline cut processing time for a 5-second clip on a single NVIDIA B200 GPU, after warm-up, from 87.26 seconds to 22.45 seconds, about 3.9 times faster. That timing covers latent upscaling, three refinement steps and 2K decoding combined.
Leiphone reports that on the RefinerBench-150 UniPercept perceptual quality benchmark the project scored 60.0, against 53.4 for the official H3-Regenerate-2K API. The benchmark score comes from the report and has not been independently verified.
Written by AI from the articles below · updated Oct 30, 5:10 AM ET
Check the sources:
Article timeline
The articles in this story. Times are ET.
- QbitAINewsNvidia and MIT release open weights for MiniMax H3-based 2K and 4K video upscaler
AINvidia and MIT's Efficient-Large-Model team released the Open-H3-Regenerate-2K model weights on Hugging Face, with inference code in the NVlabs/Sana repository. The project builds on MiniMax H3 to regenerate low-resolution video drafts at 2560×1440 (2K), with native 3840×2160 (4K) support. On a single NVIDIA B200 GPU, the team reports a 5-second clip's processing time fell from 87.26 seconds to 22.45 seconds, about 3.9 times faster.
- LeiphoneNewsNVIDIA and MIT release open 2K video upscaling project built on MiniMax H3
AINVIDIA and MIT's Efficient-Large-Model team released the Open-H3-Regenerate-2K model weights on Hugging Face, which regenerate low-resolution video drafts at 2560×1440 and up to 3840×2160 native 4K. On a single NVIDIA B200 GPU, the project's optimized pipeline cut processing time for a 5-second clip from 87.26 seconds to 22.45 seconds, about 3.9 times faster. On the RefinerBench-150 UniPercept perceptual quality benchmark, it scored 60.0 against 53.4 for the official H3-Regenerate-2K API.
Heat Trend
Not enough continuous observations to show a trend yet.