MiniMax-H3 ControlNet Union 2.0: Free AI Video ControlNet With 8 Controls (And the Apache-2.0 Alternative for US Creators)

Alibaba’s PAI team shipped MiniMax-H3-Fun-Controlnet-Union-2.0 on September 22, and it is the most capable AI video ControlNet you can download right now. One checkpoint steers MiniMax H3 with eight different control signals, from a depth map to a hand-drawn scribble, and it can inpaint a masked region of a clip on top of that.

Here is the catch you need before you spend a weekend on it: the MiniMax H3 license does not cover the United States, the EU, the UK or South Korea, and the base model wants more than one 80 GB GPU. Below: what Union 2.0 does, then the free Apache-2.0 video control stack you can run tonight on a gaming card, Wan 2.2 Fun Control in ComfyUI.

At a Glance

• What it is: a 13.5 GB “union” ControlNet for the MiniMax H3 video model. Eight controls: Canny, Depth, HED, MLSD, Pose, plus new Scribble, Layout and Gray. Video inpainting is built in.
• What’s free: the weights and the VideoX-Fun code cost nothing, and a free ZeroGPU Space already uses the model.
• What it costs in hardware: about 62 GB for the H3 transformer plus about 62 GB for the Qwen3-VL text encoder. Alibaba says that pair does not fit on one 80 GB GPU without offloading.
• The catch: the license excludes the US, EU, UK and South Korea, and that includes outputs.
• The free path for everyone else: Wan2.2-Fun-5B-Control (Apache 2.0, about 18 GB total download, native ComfyUI node).

What ControlNet Union 2.0 Adds to MiniMax H3

MiniMax H3 is the 33B omni-modal model we covered when FastVideo distilled it down to 4 steps. On its own it turns text or an image into a clip. A ControlNet lets you hand it a guide video (edges, depth, a pose skeleton) so motion and layout follow your footage.

Version 1 arrived in August with five controls. Version 2.0 changes three things, per Alibaba’s model card:

• Three new conditions. Scribble (rough sketch lines), Layout (colored bounding boxes, built with the Wan2.1-VACE annotator tools) and Gray (recolor a grayscale video).
• Twice the injection points. The control branch now attaches to 10 of the 50 transformer blocks instead of 5. Alibaba says this gives tighter structural adherence. That is the vendor’s claim, and no independent comparison has been published yet.
• Cleaner inpainting. Masked pixels now sit at mid-gray instead of near-black before encoding, which Alibaba says blends filled regions better.

Specs and the Settings That Matter

Output follows your control video. The frame count snaps down to the nearest 17 × n + 5, clips cap at 15 seconds, playback is fixed at 24 fps, and the canvas keeps your control video’s aspect ratio within a 704×1280 pixel budget.

• Use the right config. It must be minimax_h3_control_inpaint_post_norm.yaml. The v1 config loads without error but silently drops half the control weights.
• Keep guidance at 1.0. The checkpoint is guidance-distilled; higher CFG degrades output.

Audio control is off (control_apply_audio = false), so the ControlNet guides picture only. H3’s native sound still comes from the base model.

The License Problem, Plainly

The MiniMax H3 Community License grants rights only in an “Applicable Territory,” which is worldwide minus the EU, the UK, South Korea and the United States. The license defines “Output” to include results from hosted services, and it says you may not use the Materials or their Outputs outside that territory. So a US creator is not covered for local runs or for clips made through someone else’s Space.

If you are in a covered country, there are two more terms. Commercial products over US$20 million in yearly revenue need written authorization, and commercial products must show “MiniMax H3” in their UI. I’m not a lawyer, and this isn’t legal advice. Read the license yourself before you ship client work.

AI Video ControlNet Comparison: MiniMax H3 vs. the Free Alternatives

Try It Free: Video Control You Can Run Tonight

Option 1: Wan2.2-Fun-5B-Control in ComfyUI (Apache 2.0, works everywhere). This is also from Alibaba PAI, the same team behind ControlNet Union 2.0. It supports Canny, Depth, Pose, MLSD and trajectory control, and it was trained on 121-frame clips at 24 fps. ComfyUI runs it natively with the Wan22FunControlToVideo node.

1. Install ComfyUI: git clone https://github.com/comfyanonymous/ComfyUI.git && cd ComfyUI && pip install -r requirements.txt, then python main.py.
2. Download three files from Comfy-Org’s Wan 2.2 repack: wan2.2_fun_control_5B_bf16.safetensors (10 GB) goes in models/diffusion_models, umt5_xxl_fp8_e4m3fn_scaled.safetensors (6.7 GB) goes in models/text_encoders, and wan2.2_vae.safetensors (1.4 GB) goes in models/vae.
3. Open Workflow → Browse Templates → Video → Wan2.2 Fun Control. The template is built for the 14B model, so swap the two diffusion loaders for a single loader pointing at the 5B file, and point the VAE loader at wan2.2_vae. The node detects the 5B model’s 48-channel latent by itself.
4. Canny preprocessing is built in. For depth and pose, install comfyui_controlnet_aux by running git clone in custom_nodes.

If you have a 24 GB card, the official 14B tutorial gives real numbers. On an RTX 4090D at 640×640 and 81 frames, the fp8 models took about 524 seconds. With the 4-step lightx2v LoRAs that dropped to about 79 seconds on a warm run. SageAttention can cut that further.

Option 2: no GPU at all. The free Wan 2.1 VACE Fast Space runs Wan2.1-VACE-14B at 6 steps on ZeroGPU. It isn’t pose or depth control; it offers reference-image, first/last-frame and transition modes. Expect a daily GPU quota and queues.

Option 3: inside the licensed territory with a big GPU. Clone VideoX-Fun, put MiniMax-H3 and the Union 2.0 checkpoint in models/Diffusion_Transformer, set the paths at the top of examples/minimax_h3_fun/predict_v2v_control.py, and run python examples/minimax_h3_fun/predict_v2v_control.py. On a single 80 GB GPU, use model_group_offload. The RioShiina MiniMax-H3 Space is listed as using this checkpoint on free ZeroGPU.

Honest Trade-Offs

• No independent quality numbers yet. The “tighter adherence” claim comes from Alibaba’s own samples.
• Wan 5B isn’t H3. It has no native audio, it’s lower fidelity than a 33B model, and it doesn’t have Scribble, Layout or Gray control. You’re trading ceiling for a license you can actually use.
• The 8 GB figure is a floor, not a comfort zone. Offloading works but is slow; budget plenty of system RAM.
• LTX is cheaper to download, not to run. The union IC-LoRA is only 650 MB, but it needs the 22B LTX-2.3 base, and companies earning $10M or more a year need a paid license.

Workflow Tips

• Shoot your own reference on a phone. A 5-second clip of you doing the move, run through DWPose, beats any stock clip, and you own it.
• Trim the control clip before you generate. H3 snaps to 17n+5 frames and Wan 2.2 5B was trained on 121 frames. Cut the source to fit so you don’t lose the ending.
• Match the control to the job. Use depth for camera moves through a scene, pose for dance or acting, and Canny for keeping products and logos rigid.
• Describe the camera in the prompt. Alibaba says detailed prompts covering scene, subject and camera give more stable results.
• Prototype at 480p. Lock the motion at low resolution, then upscale the keeper. If you want a faster draft model for that step, see our SANA-Video 2.0 4-step write-up.

To see how the H3 base model handles continuity, watch our review MiniMax H3 Video Extend: Can It Keep the Scene Going? on the Creative AI Show YouTube channel.

FAQ

Is MiniMax-H3-Fun-Controlnet-Union-2.0 free?

Yes, the weights and code are free to download. Use is licensed only outside the US, EU, UK and South Korea, and large commercial users (over $20M a year in revenue) need written authorization.

Can I use MiniMax H3 ControlNet in the US?

Not under the community license. The US is an Excluded Territory, and that covers outputs, including those made through hosted services. Wan2.2-Fun-Control (Apache 2.0) is the closest unrestricted alternative.

What are the system requirements for MiniMax H3 ControlNet Union 2.0?

You need about 138 GB of weights: a 13.5 GB control branch, a ~62 GB transformer and a ~62 GB text encoder. It also needs an 80 GB GPU running with model_group_offload or qfloat8 offload. It’s not a consumer-card model.

What is the best free AI video ControlNet for 8 GB of VRAM?

Wan2.2-Fun-5B-Control in ComfyUI. It’s Apache 2.0, about 18 GB, and runs with ComfyUI’s native offloading, slowly.

Accurate as of September 27, 2026. Licenses, file sizes and Space availability change fast.