GitHub radar
Better faces and HDR lighting in MiniMax-H3 video
A community contributor spent three days fine-tuning MiniMax AI's H3 video model and released a version that fixes facial distortion, removes unnatural skin shine, and adds HDR color response.
Minimax-h3_Singularity is a community fine-tune of MiniMax-H3, the video generation model from MiniMax AI. The author WarmBloodAban describes it as a comprehensive fine-tuned fusion model built to fix specific visual weaknesses in the base version. According to the model card, the work targeted facial distortion in medium and long shots, unnatural skin shine, HDR color range, dynamic motion, and VFX rendering. The author reports spending three days on precise model pruning and weight optimization. The model is fully compatible with ComfyUI and supports text-to-video, image-to-video, reference-to-video, and video-to-video workflows. The card states that the base model's prompt adherence and style adaptability are preserved in full.
Why a vibe-coder should care
If you already use MiniMax-H3 in ComfyUI for video generation, this version addresses the most visible quality issues without changing your existing setup. Facial quality in wider shots is a common complaint with the base model, and this fine-tune targets it directly. That said, this is a community project, not an official MiniMax release — treat it as an experiment worth trying, not a guaranteed improvement.
How to install
Copy this and send it to your agent — Claude Code, Codex, any of them:
Install this MiniMax-H3 variant in ComfyUI in place of the base checkpoint: https://huggingface.co/WarmBloodAban/Minimax-h3_Singularity — load the model and open my existing video generation workflow so I can compare the output quality against the original.
Requires a GPU with at least 16 GB of VRAM — the same as standard MiniMax-H3. A regular laptop will not run it.
Open on Hugging Face▌ More finds