Player not loading? Watch on YouTube
This tutorial tests Qwen Image 2.1 Fun ControlNet Union in ComfyUI. The 7.55 GB model file goes in models/model_patches and supports eight control types plus inpainting. At recording time, the Apply Qwen Image 2.1 Fun ControlNet node required Kijai's pull request version rather than the normal ComfyUI release; the speaker warns that other versions may show a missing node.
Pose tests expose two problems. A vertical character image and horizontal pose reference produce a mismatch. The speaker reports better results after resizing both to the same dimensions. An optional upscale step uses their 32 GB of RAM; they advise skipping it with less RAM. A difficult lunge still produces a twisted hand after resizing. Describing the pose in the prompt and reducing strength from 1 to 0.8 fixes that example. The speaker prefers DWPose over depth processing for pose control.
Depth Anything V2 guides living-room edits into a dark nighttime interior and an abandoned room. The speaker reports that furniture placement stays consistent. A separate scribble experiment bypasses ControlNet and uses the drawing as a reference image, producing a realistic image and a zombie transformation. Inpainting examples replace a woman's top with a T-shirt and a fruit bowl with a glass vase of white flowers. These are individual test results, not guarantees for every input.
The source description links the creator’s workflow on Patreon. The Union ControlNet is an Alibaba PAI add-on for Qwen Image 2.1; it is not presented here as a separate release of the base Qwen model.
