AI_ModelScopeImageEdit

Qwen-Image-Edit is one of the genuinely good open models to come out of Alibaba's Tongyi labs - it's the instruction-based editor that took over a lot of jobs that used to need a mask plus a ControlNet plus a face adapter stacked together: object removal, garment swaps, relighting, re-posing, camera angle changes, all from a picture and a sentence. It won that role over Flux Kontext largely on licensing - Apache 2.0 against a license that made NSFW LoRAs a violation - and the community built an entire LoRA ecosystem on top of it as a result. AI_ModelScopeImageEdit is a hosted-API route to that model through ModelScope, so you get it without downloading the weights or having the VRAM to run it locally.

Like its sibling AI_ModelScope_image, the input labels ship in Chinese: 预设模型 (preset model, default Qwen/Qwen-Image-Edit, with MusePublic/Qwen-Image-Edit as an alternate hosted variant), 提示词 (your edit instruction - the sentence describing the change), 图像宽度/图像高度 (output width/height, 64–8192 in steps of 64, default 1024×1024), 张数 (how many output variations, 1–4), 超时时间 (timeout, 1–900 seconds, default 300 - meaningfully longer than the chat nodes' defaults, which tracks with image editing being a slower job than a text reply), seed, MODELSCOPE_API_TOKEN, and 自定义模型 for a free-text override. You can feed it up to five source/reference images (图像1 through 图像5), which is generous - enough room for a source photo plus several reference images if your edit needs style or composition guidance from more than one input. Outputs: 图像 (the edited IMAGE), 图片链接 (a direct URL to the result - handy if you want to grab it outside ComfyUI too), and status.

Worth knowing going in, and this comes straight from how Qwen-Image-Edit itself behaves, not a quirk of this specific node: the model re-emits the whole frame on every edit rather than touching only the region you meant to change, so pixels you didn't ask about can shift, and running the same image through multiple edit passes in a row tends to let faces and fine detail drift a little more each time. Keep edit chains short, and if you need surgical precision on one small region, a traditional mask-based inpaint workflow is still going to beat an instruction editor.

Installing it: ComfyUI Manager, search "ComfyUI-Apt_Preset," or clone manually - cd ComfyUI/custom_nodes && git clone https://github.com/cardenluo/ComfyUI-Apt_Preset - then install.bat and restart.

Getting a token: sign up at modelscope.cn and pull an API token from your profile settings. ModelScope's free tier has real usable quota for a lot of models, which makes it a reasonable way to try Qwen-Image-Edit before committing to running it locally.

Troubleshooting: a failed or empty call is almost always MODELSCOPE_API_TOKEN - verify it's set and not expired. If a batch of 张数 > 1 is timing out, raise 超时时间 before assuming something's broken; multiple output images at high resolution genuinely take longer than the 300s default budgets for. And if an edit changed more of the image than you asked for, that's expected model behavior rather than a bug in this node - see the whole-frame re-emission note above, and consider running a lower-strength or more targeted pass instead of one big instruction.