Relight · Edit

Relighting is one of the few compositing problems a purpose-built model solved cleanly, and then the whole category got eaten by instruction editors. IC-Light was the answer in 2024 - the KB's writeup of it is blunt that its last code commit was the day after release, and that people didn't stop relighting, they stopped using IC-Light to do it. What replaced it was exactly this: "relight the image with warm sunset light from a window on the left", typed into an editor model.

Relight · Edit is that sentence, one node, 4.5 credits an image.

What it is

popnow/relight-v1/edit behind a ComfyUI node in the PopNow pack. Your photo goes up to PopNow's servers, comes back relit as a normal IMAGE. No IC-Light checkpoint, no Lighting Preference, no mask-and-composite pass. If you want the local version of this idea in 2026 you're downloading a relight LoRA for Qwen-Image-Edit and wiring it into an edit graph; this node is the hosted shortcut for when you have one photo and one deadline.

The trade is the same as every API node in this pack, and worth saying once plainly: cheap, no VRAM, and your picture is on someone else's server. For a stock photo or a render you made yourself, fine. For a client's unreleased product shot, think about it first.

The one input that matters

prompt is the new light, and unlike most prompt fields in this pack it has a shape you should follow. The tooltip's own example is the template: "Relight the image with warm golden sunset light from a window on the left." That's four properties in one sentence - source (window, softbox, fire, overcast sky), direction (left, right, above, behind), colour (warm golden, cold blue, neutral) and hardness (soft, harsh, diffused).

Write all four and you get predictable results. Write "make it dramatic" and you get a coin flip, because the model has to guess which of those four axes you meant.

image_url is the source photo - up to 10 MB, and sharpness matters. A blurry source gives the model fewer cues about the existing geometry of the scene, and a relight that misreads geometry produces the classic tell: shadows in two directions at once.

aspect_ratio defaults to model default, keeping your framing. Leave it.

output_format is JPEG; if you're compositing the result back into a matte, switch it to PNG. seed starts fixed so a repeat queue reuses the run rather than billing you again; randomize control-after-generate to explore at 4.5 credits per take.

Outputs: images, task id.

Install

Manager → search PopNow → Install → restart, or:

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

Then a key from popnow.ai/api/console/keys via Settings > PopNow, or:

POPNOW_API_KEY=sk-popnow-... python main.py

No dependencies, no model files, no GPU. Restart once more after the key is set - node types are generated from your account's catalog at startup, which is also how new PopNow models turn up without a pack update. Remove the old single-file popnow_api.py if it's still there.

Where people get burned

A relight is a redraw. The model doesn't move a light around your scene; it re-renders the whole frame under a new description. Skin, fabric and hair will change subtly, and text, logos and fine pattern will not survive. Check anything with lettering before you ship it.

Consistency across a set is the hard part. Two photos relit with the same prompt will not match each other exactly, because each one re-reads its own scene. Batch product shots relit one by one end up looking like they came from two different shoots. The fix is to relight one hero shot and composite the rest, not to relight everything.

Ambiguous existing light fights you. A source with strong mixed lighting - window on one side, tungsten on the other - gives the model two stories to reconcile. Cleaner sources relight better.

Spend. Every randomize is a billed run, a failed run is refunded, and a run you cancel from the ComfyUI queue keeps going on PopNow and is charged if it succeeds. Set Max credits per run in Settings > PopNow while you're finding out what the model does with your prompts.