Modelverse Qwen Image Edit

Alibaba's Qwen-Image-Edit has a reputation in the image-editing world for doing what you ask, and this node is that model with a ComfyUI face. Same deal as everything in this pack: it's not running locally - it's a call into UCloud's Modelverse API, which means no model download, no LoRA fiddling, no VRAM concerns, and a meter running somewhere in the cloud. For one-shot style flips and targeted edits, it's the node in the pack most likely to make you say "huh, it actually did it."

What it is

A single-image editor. Feed it an IMAGE tensor and an instruction like "convert to quick pencil sketch" (the shipped default) and it re-renders the image to match. Where a typical img2img pass needs you to tune denoise and hope, Qwen-Image-Edit understands natural-language instructions - "make it look like a watercolor," "turn the background black and white," "put a hat on him" - and follows them.

The inputs that matter

• client - required wire from the Modelverse Client / Secret Client node.
• image - required. The source image to edit.
• prompt - the instruction. This is the star input; be specific and imperative.
• negative_prompt - things to avoid. Worth filling once you know what the model keeps sneaking in.
• strength - default 0.8, the img2img transformation strength (0–1). Crank it for big restyles, drop it for subtle touch-ups. This is the main dial you'll touch.
• steps - 20 default, up to 50. More steps, more faithful, slower (and billed).
• guidance_scale - 2.5 default, 1–10. Higher = sticks closer to the prompt.
• num_images (1–4) and num_requests (1–10) - batch knobs. This is where your bill grows fastest; resist the urge.
• seed - -1 for random, or set it to reproduce a result.

Output is a single image (IMAGE tensor), ready to preview or save with any standard node - no extra download step like the video nodes need.

Installing and keying up

Standard for this pack. Fastest path: ComfyUI Manager → search "ComfyUl-UCloud" → Install → restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/ucloud/comfyui-plugin

restart ComfyUI, done. Dependencies (requests, pillow, numpy, openai, GitPython) are auto-handled by Manager; there are no model files to place. The only real setup is the ModelVerse API key - from the UCloud console, entered via the Modelverse Secret Client node (stores the key in secrets.json under a name you type, keeping it out of the workflow) or a config.ini MODELVERSE_API_KEY=... line.

Common issues

• Edits come back barely changed - strength too low. Bump it; the default 0.8 is a starting point, not a law.
• "API key is not set" - the client node is empty. It's always the client node.
• Edit quality degrades at high resolution - Qwen-Image-Edit is good but the API re-encodes your image; very large inputs can come back soft. Downscale before sending, edit, upscale after.
• It restyles the whole image when you only meant one object - the model has no mask input in this node. For surgical edits, crop to the region, edit, then composite back.

The one-liner: describe the change in plain words, don't hunt for the right CFG. It's a different editing paradigm than diffusion-strength roulette, and once you're used to it, it's hard to go back.