Qwen Image 2512 LoRA · Edit

Single-image editing handles "make the jacket green". It does not handle "put the person from photo A into the room from photo B wearing the coat from photo C", which is the actual job a lot of people have. That's what this node is for: one to three source images in, one composite out, with your own LoRA stack on top.

It's built on the 2512 revision of Qwen-Image-Edit - the late-2025 refresh of the same 20B family - and priced at 4.5 credits ($0.045) per image.

The naming convention is the whole trick

The prompt tooltip says it outright: "Refer to the sources as Picture 1, Picture 2 and Picture 3." Not "the first image". Not "the woman". Capitals matter less than consistency, but the convention does not - this is how the model knows which of your three uploads you're talking about. Write prompts like "Picture 1 standing in the kitchen from Picture 2, wearing the jacket from Picture 3".

And note the second tooltip, the one that saves arguments: "The result follows the first one." Picture 1 sets the output's framing and shape. If your composite comes back with the wrong crop, you connected the images in the wrong order, not the wrong resolution - reorder the sockets rather than fiddling with settings.

What you get to set

prompt - the instruction, with Picture N references and, if your LoRA has a trigger phrase, the phrase first.

image_urls - an Autogrow group: one socket to start, and a new one appears each time you connect something, up to three. Each connection is uploaded separately and sent as an ordered list. A whole IMAGE batch on one socket counts as several pictures, which is a fast way to confuse yourself - prefer one Load Image per socket when order matters.

aspect_ratio - model default, i.e. the field is left out of the request so the result follows the first source. Override it only when you want a deliberate reframe.

loras - optional, up to three PopNow LoRA nodes, each with a scale. Point source at a .safetensors in ComfyUI/models/loras, a Hugging Face repo, a Civitai page, or a URL. Local files upload once and get reused for twelve hours; gated repos read HF_TOKEN/CIVITAI_API_KEY from the environment.

output_format - JPEG by default, PNG if the result is going back into another edit.

seed - fixed by default. Re-queueing the same graph returns the same cached image rather than spending another 4.5 credits. Flip control after generate to randomize for a new take.

Outputs: images, and task id for the API console.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/popnow-ai/popnow-comfyui

Restart ComfyUI. Manager route is the same thing with fewer keystrokes: search PopNow, Install. There's nothing to pip install - the pack declares zero dependencies beyond ComfyUI's own bundled libraries. You'll need Python 3.10+ and a ComfyUI with the V3 node API.

Then popnow.ai/api/console/keys → Settings → PopNow → paste → restart once more so the catalog loads from your account. Headless? POPNOW_API_KEY in the environment beats the settings field, and keeps the key out of any workflow you hand to someone else.

Reality checks

Qwen-Image-Edit re-emits the entire frame, so with three references it has more chances to drift. Expect the unmentioned parts of Picture 1 to shift slightly. Stacking two edits is fine; stacking five and expecting pixel-stable faces is not - that's the model's documented weakness, not your prompt.

The 2512 revision has its own LoRA library and it is not interchangeable with the 2509 or 2.1 sets. A LoRA trained for Qwen-Image-Edit-2509 will load and produce something, but "something" is the honest description. Match the LoRA to the model family and the render is stable.

Codes you'll meet: 40201 no credits, 400 naming a rejected field, 40004 an unservable combination, and refunded content_rejected/generation_failed runs. Paste task id into support if a run ends with a request_id.