GitHub radar
Turn MiniMax-H3 video drafts into 2K for free, and faster
This is a free, open replacement for MiniMax-H3's paid upscaling step: it turns a low-resolution draft into a sharp 2560×1440 clip, noticeably faster than the official version.
A ready-made package — model tweaks plus a separate decoder (the part that turns data into a picture) — that adds detail to MiniMax-H3 video and raises its resolution from 768P to 2560×1440.
When it helps
Useful if you already run a MiniMax-H3 video pipeline and need an upscaling step; skip it if you don't work with MiniMax-H3 or video generation.
Pros
- Free replacement for a paid upscaling feature
- Speeds up the final step almost 4x on strong hardware
- Lightweight decoder makes its part nearly 10x faster
- Output matches the original per the reported checks
Cons
- Needs a server-grade NVIDIA B200 GPU, no home setup
- License is pieced together from several sources, check each one
- Only works as a step inside a MiniMax-H3 pipeline, not alone
How to set it up — step by step
- 1Open the Open-H3-Regenerate-2K page on Hugging Face: huggingface.co/Efficient-Large-Model/Open-H3-Regenerate-2K.
- 2Send your agent the text from the block below to break down the package, hardware, and license requirements.
- 3Check with your agent whether you have access to a server-grade NVIDIA B200 GPU.
- 4Ask your agent how to plug this step into your own MiniMax-H3 pipeline.
- 5Before commercial use, review each of the several licenses separately.
Text for your agent
Copy this and send it to your agent — Claude Code, Codex, any of them:
Open the Open-H3-Regenerate-2K model card on Hugging Face: https://huggingface.co/Efficient-Large-Model/Open-H3-Regenerate-2K — explain what's in the package, what hardware and licenses it needs, and how to plug it into a MiniMax-H3 pipeline.
A server-only tool — not something you run at home. The authors' own benchmarks use a single NVIDIA B200 GPU, a datacenter-class card, not a gaming or consumer one.
Open on Hugging Face▌ More finds