qwen / 20b / editImage

The instruction-edit node for Qwen-Image, running on Civitai's sdcpp fleet. Qwen-Image-Edit is the model that made masks and adapters optional for a whole class of edits - you hand it a picture and a sentence ("remove the watermark", "change the sign text to 'Bella's Books'", "make the sky sunset") and it does it, re-emitting the whole frame. This node wraps that for the cloud, so the edit happens without a 20B model on your disk and without the two-encoder setup Qwen needs locally.

Its specific edge over the Flux edit nodes is text. If your edit involves words - fixing a misspelling on a storefront, changing a label, replacing logo copy - Qwen's typography is the best in the open world, and it survives the edit. That's the workflow this node exists for: images where the text is the point, and the edit is about changing it.

How it works

Standard pack client: submit engine: sdcpp, ecosystem: qwen, model: 20b, operation: editImage with your images, prompt, and required output size; long-poll; get the edited IMAGE back.

The inputs that matter

• prompt (required) - the edit instruction, plain language. "Change the text on the menu board to read X" is exactly its home turf.
• images (required, IMAGE) - the source. Unlike the Flux edit nodes, this one requires it - there's no accidental "edit with no image" here.
• width / height (required) - the output size, no inference from the input. Set it to match your source.
• version - latest / 2509 / 2512. The 2511/2512 builds are the current open editors.
• steps / cfg_scale - 20 / 2.5 defaults.

Install

Pack install: ComfyUI Manager → search "Civitai Comfy Nodes" (publisher civitai) → Install → restart. Or clone https://github.com/civitai/civitai-comfy-nodes.git into custom_nodes, pip install -r civitai-comfy-nodes/requirements.txt (just requests), restart.

Common issues

Credentials: Civitai Auth node into api_config or CIVITAI_API_TOKEN (from civitai.com/user/account); headless installs need the env var.

The structural weakness to plan around: Qwen re-emits the whole frame, so pixels you didn't ask about move a little - faces drift across a chain of edits, backgrounds shift. It's the model's known design, not a failure. If an edit must be surgical, the community's standard answer is to bolt a mask on around the region you care about.

And the required width/height: it's easy to wire an image, type a prompt, and queue with blank size widgets - the node catches it locally and blocks. Set the size, check the text, and remember each edit is a billable job.