Introduction

MiniMax H3 already sings video with synced audio — the missing desk muscle for a lot of creators has been structural steering. Tonight Alibaba PAI’s VideoX-Fun crew ships MiniMax-H3-Fun-Controlnet-Union-2.0: one control branch that eats eight condition videos, injects twice as densely as v1, and cleans up masked inpaint blending. Same-day, the VideoX-Fun repo lands the matching configs and predict scripts.

What’s new in 2.0

Think of v1 as five control dials (Canny, Depth, HED, MLSD, Pose). 2.0 adds Scribble, Layout, and Gray — freehand sketches, color-coded bounding-box layouts (Wan2.1-VACE style), and luminance videos — so you can steer without baking a perfect edge map every time.

Under the hood the control branch grows from 5 blocks to 10, skipping into transformer layers 0, 5, 10, …, 45 instead of only five sparse attachment points. That denser skip-add is why the card claims tighter structural adherence at control_context_scale = 1.0.

Inpaint also changes recipe: v1’s pre_norm holes sat near extreme dark in VAE space; 2.0’s post_norm zeros masked pixels after ImageNet normalization (mid-gray holes), following Wan 2.1’s cleaner blend. Checkpoint size roughly doubles to ~13.5 GB of control-only weights you load on top of base MiniMax-H3.

Union v1Union 2.0

Conditions

5

8 (+ Scribble, Layout, Gray)

Control blocks

5

10

Inpaint holes

pre_norm (~−2)

post_norm (0 / mid-gray)

Control file

~6.8 GB

~13.5 GB

Config

minimax_h3_control.yaml

minimax_h3_control_inpaint_post_norm.yaml

Guidance-distilled still: keep guidance_scale = 1.0. Raising CFG double-applies guidance and muddies the shot. Frame count snaps to the familiar 17 * n + 5 VAE grid, canvas follows the control video’s aspect at a 32-multiple budget, 24 fps, duration capped around 15 s.

Where to get it

Quick VideoX-Fun path

git clone https://github.com/aigc-apps/VideoX-Fun.git
cd VideoX-Fun
mkdir -p models/Diffusion_Transformer
# place MiniMax-H3 under models/Diffusion_Transformer/MiniMax-H3
# place 2.0 safetensors under .../MiniMax-H3-Fun-Controlnet-Union-2.0/
python examples/minimax_h3_fun/predict_v2v_control.py
# or with mask:
python examples/minimax_h3_fun/predict_v2v_control_inpaint.py

Point config_path at config/minimax_h3/minimax_h3_control_inpaint_post_norm.yaml. Loading the old five-block yaml against 2.0 is a silent failure — half the control weights drop under strict=False and outputs look “almost right” for the wrong reason.

Memory reality check from the card: full transformer (62 GB) plus Qwen3-VL text encoder (62 GB) will not sit happily on one 80 GB GPU fully resident — use model_group_offload or model_cpu_offload_and_qfloat8.

ComfyUI note

Community ComfyUI-H3-FunControl graphs already chase the Fun Control lane for depth/pose/canny-style control on pruned H3 bases (often with Kijai curve-form checkpoints). Treat official Alibaba ~13.5 GB Union-2.0 as the VideoX-Fun / full-width path first; don’t assume a pruned Comfy loader will swallow the PAI file without a matching custom node update. Watch the Comfy issue threads if you live in Manager-land.

License

Derivative of MiniMax-H3 under the MiniMax H3 Community License — read territorial restrictions and Acceptable Use before you ship anything. Same family constraints you already know from HyperFlow, H3-LongVideos, and the OmniVBench coverage.

Original Source

https://huggingface.co/alibaba-pai/MiniMax-H3-Fun-Controlnet-Union-2.0

https://github.com/aigc-apps/VideoX-Fun/commit/2beb171099e932faf95030b604fb44e04e961276

https://github.com/aigc-apps/VideoX-Fun

Conclusion

Eight conditions, denser skips, and a saner inpaint hole recipe make Union-2.0 the first H3 Fun Control drop that feels like a full desk toolkit rather than a five-dial demo. Pair base MiniMax-H3 with the new VideoX-Fun configs, keep CFG at 1.0, and let Scribble or Layout do the heavy steering when Canny is too brittle.

Try MiniMax H3 on Gen → https://artrealmai.com/gen?utm_source=magazine&utm_campaign=gen&utm_content=h3-fun-controlnet-union-2-0

—Aurelia ♡