GJuarez67/MiniMax-H3-Fun-Controlnet-Union-Demo

MiniMax-H3-Fun-Controlnet-Union — Hugging Face Space

A showcase Space for the MiniMax-H3-Fun-Controlnet-Union model by alibaba-pai.

What is this?

MiniMax-H3-Fun-Controlnet-Union is a ControlNet-Union checkpoint for the MiniMax-H3 video diffusion transformer. One single checkpoint handles Canny, Depth, HED, MLSD, and Pose control conditions for video-to-video generation — no per-condition checkpoint switching — and also supports video inpainting.

The control branch (~6.8 GB) is loaded on top of the base MiniMax-H3 transformer and injects control signals at 5 points (layers 0, 10, 20, 30, 40) of the 50-block architecture.

Space contents

This Space shows:

• Model overview — description, file table, and key features
• Inference defaults — the settings used for all published results
• 5 result pairs — control input + generated output for each control type (Canny, Depth, HED, MLSD, Pose), sourced directly from the model repo
• Setup guide — how to run inference via the VideoX-Fun pipeline

Inference requirements

Running the actual model requires:

• The base MiniMax-H3 transformer (~62 GB)
• The Qwen3-VL text encoder (~62 GB)
• This ControlNet-Union checkpoint (~6.8 GB)
• VideoX-Fun installed with the matching minimax_h3_control.yaml config

The full model does not fit a single 80 GB GPU fully loaded — use model_group_offload or model_cpu_offload_and_qfloat8.

See the Space app for the full step-by-step inference guide.