Qwen-Image ControlNet img2img ⚡

Qwen-Image's ControlNet story is a mess of formats, and this node is the broom. The ecosystem circles three different attachment mechanisms - InstantX/Union checkpoints, Qwen-Image-Fun, and DiffSynth patches - and each wires into ComfyUI differently: some attach to conditioning like a classic ControlNet, some patch the model directly. The pack's Qwen-Image ControlNet Loader figures out which format you loaded and hands you a QWEN_IMAGE_CONTROL token. This node does the attaching, so you never have to know or care which mechanism your file needs.

It's everything Qwen-Image img2img does - prompt encoding, VAE-encoding your source photo, building the init latent - plus control attachment, plus (when connected) a mask for inpaint checkpoints. For DiffSynth patches it clones the MODEL and applies the patch; for InstantX/Union/Fun it stamps the control hint onto your CONDITIONING the same way stock ControlNetApplyAdvanced does. Outputs are model, positive, negative, latent, denoise straight into a stock KSampler.

The inputs that matter

• qwen_control - the loaded checkpoint token from Qwen-Image ControlNet Loader. Ignored without it.
• images - your init photo(s) for img2img, and in the auto modes the source the control map is derived from. Leave unconnected for txt2img.
• control_mode - picks how control_image is produced, since nothing in the file says what preprocess it wants: manual (default, connect your own map - required for inpaint), auto_canny (plain cv2 edges, no download), auto_depth (Depth Anything V2, auto-downloads weights), or none (skip control for this call even if qwen_control is connected). Explicit control_image always overrides.
• control_strength (default 1.0) - how hard the condition bites.
• mask - for inpaint checkpoints, the region to fix. Note the tooltip truth: mask only refines the region, it doesn't replace the control image - you still need control_image even for inpaint.

Installing

cd ComfyUI/custom_nodes
git clone https://github.com/ChrisColeTech/ComfyUI-GGUF-Loader
pip install --upgrade gguf

Restart, drop your Qwen-Image model (GGUF or safetensors), Qwen3 encoder, VAE and ControlNet checkpoint in the usual folders. The nodes sit under 🤖 CCTech/Qwen-Image. Preprocessors beyond canny/depth (normal, soft edge, pose) come from the separate ComfyUI-ControlNet-Nodes package - wire their output into control_image with control_mode=manual.

Gotchas

Same guard rails as its Krea2 sibling: qwen_control with nothing usable to attach (outside none mode) raises rather than sampling a half-configured model, and control_image with no qwen_control is ignored with a warning. Qwen-Image spans more control types than the pack can auto-detect from the file, so the auto modes are genuinely best-effort - for a pose checkpoint, manual with a real preprocessor is the reliable path.