P-Image LoRA · Edit

Multi-image editing - "put the jacket from picture 2 on the person in picture 1, in the room from picture 3" - is the feature Qwen-Image-Edit made mainstream, and every hosted service has a version of it now. What varies is the price and how many references you get.

P-Image LoRA · Edit is the cheap end: 1.8 credits an image, up to five source images, one LoRA of your own. That's under two cents for a composition job that would cost you a 20 GB checkpoint download to do locally.

What it is

A cloud node over pruna-ai/p-image-lora/edit in the PopNow pack. No VRAM, no weights, no sampling on your machine - your images are uploaded, the edit runs on PopNow's servers, and the result comes back as a normal IMAGE. The API-node writeup in the KB is the honest framing of this whole category: you're trading per-call cost and an upload of your inputs for hardware you don't have to own.

"P-Image" is Pruna's fast image model, and the node's own summary sells it as exactly that - fast renders. For this node, that's the pitch: multi-image edits priced low enough to iterate.

The interesting input: image_urls

image_urls is an Autogrow socket group, which is the pack's tidy way of expressing "one to five images". You wire one image and it grows another socket. Connect five and all five get uploaded.

Two things follow. First, the prompt has to reference them by number - the tooltip says to refer to them as image 1, image 2 and so on, and it means it literally. Vague pronoun-ish prompts ("put the jacket on her") work worse than "place the jacket from image 2 on the person in image 1".

Second, the first image is the layout. Without an aspect_ratio set, the result follows the first one. So the order you connect them in is a decision, not alphabetisation.

prompt is "what to change and what to keep", and that second half is worth typing. Edit models drift; saying "keep the pose and the background" is cheap insurance.

aspect_ratio defaults to model default, which means inherit from image 1. Options are the usual 1:1 / 16:9 / 9:16 / 4:3 / 3:4 / 3:2 / 2:3 - note there's no custom here, unlike the text-to-image sibling. If you need exact pixels, you're out of luck on this node.

output_format defaults to jpeg; use png for anything compositing onward.

seed starts fixed so a repeat queue doesn't bill you twice.

loras takes one LoRA - not three, like the Qwen nodes in this pack. Add a PopNow LoRA node, point source at a .safetensors in ComfyUI/models/loras, a Hugging Face repo, a Civitai page or a direct link, set scale, wire it in. Gated files need HF_TOKEN or CIVITAI_API_KEY in the environment; the pack won't take tokens from a widget, because widgets get saved into every workflow you share.

Outputs: images and task id.

Install

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

Or Manager → search PopNow → Install → restart, then a key from popnow.ai/api/console/keys in Settings > PopNow (or POPNOW_API_KEY=sk-popnow-... python main.py). Zero dependencies, zero model downloads.

Where people get burned

Five images is five uploads. The pack caches uploads by content hash for twelve hours, so re-running with a different prompt doesn't re-send the same pictures - but the first run of a five-image graph is a real upload, and each image is re-encoded to PNG at up to 10 MiB.

image 1 means socket one. Autogrow groups order by socket index, and "first" is whichever socket you wired first. People connect things in a hurry and then wonder why the swap went the wrong way.

A LoRA trained for a different base will not help here. One slot, and it wants a LoRA trained for P-Image editing. Feeding it a Qwen or Flux LoRA costs you a run and produces either an error or mush.

The cheapest edit isn't the right edit. At 1.8 credits you can afford forty attempts, which is often what a fiddly multi-image composition actually takes - so budget for iterations rather than one perfect prompt, and set a spend ceiling in Settings > PopNow so an afternoon of experimenting doesn't surprise you.