Qwen Image In Context Control Union. Adds structure control on top of Qwen-Image. Works with canny, depth, lineart, softedge, normal, and openpose. Simple setup. Clear outputs.
Clean LoRA on Qwen-Image. Uses In-Context conditioning, not hard wiring. Trains in DiffSynth-Studio with the Qwen-Image-Self-Generated-Dataset. Prompts usually start with “Context_Control.” so the model switches into the right mode.
Qwen Image In Context Control Union basics in normal words
Base is Qwen-Image.
Adapter is LoRA for structure.
Method is In-Context conditioning.
Controls include canny / depth / lineart / softedge / normal / openpose.
Framework is DiffSynth-Studio.
Data is Qwen-Image-Self-Generated-Dataset.
Prompt tip: begin with “Context_Control.”
What to expect from the control types
Canny, depth, lineart, softedge, and normal give tight shape guidance. Good “hold the form” feeling.
Openpose is different. Due to how pose maps work, it won’t do perfect point-to-point matching like the others. It gets close, not exact. That is normal.
Where everything lives (kept simple)
Code sits in the DiffSynth-Studio repo on GitHub. The Qwen-Image base and annotators are on the usual hubs (ModelScope / Hugging Face). The LoRA is on the model page—grab model.safetensors from Hugging Face here: Qwen Image In Context Control Union model.safetensors.
