How to Split an Image into Layers

Last month a client sent me a product shot and asked for "the same image, but winter." The file was a flattened JPG. No PSD, no layers, nothing. Five years ago that request meant an hour of careful masking, or an awkward conversation about why it wasn't possible.

This is the problem layer decomposition actually solves, and it's worth understanding what it does before you rely on it.

What layer decomposition is

You upload a finished image — a poster, a product photo, an AI generation, whatever. A model (we use Qwen Image Layered) looks at the picture and rebuilds it as a stack of transparent layers instead of one flat bitmap. The subject lands on one layer. The background on another. Text, shadows, and small objects each get their own.

The output isn't a guess with a checkerboard behind it. Each layer is a real PNG with real transparency, positioned where it was in the original. Stack them back up and you get the source image again, pixel for pixel.

Doing it, step by step

Open the editor and upload your image. JPG, PNG, WEBP, and AVIF all work, up to 25 MB.

Pick a layer count. Four to eight layers is the sweet spot for most images. Fewer and everything gets glued together into mega-layers; more and you spend your time hunting through the stack for the thing you wanted.

Hit decompose and wait. On a typical product shot it takes under a minute. You'll see the layer stack fill in on the left — usually named something like Main Subject, Text / Graphics, Background, Shadow.

That's the whole process. What you do next is where it gets interesting: hide a layer to see what's behind it, drag one out of place, or edit a single object while the rest of the image stays exactly as it was.

What comes out the other end

The honest version:

• Clean separations: product on plain background, text over flat areas, distinct objects. These come out as genuinely usable layers.
• Acceptable separations: soft shadows, semi-transparent effects. You get a layer, but the edges are soft and sometimes you'll want to clean them up.
• Messy separations: hair against busy backgrounds, elements baked together by heavy JPEG compression, reflections fused to glass. The model tries. Sometimes the result is good enough to edit around; sometimes you're better off leaving that region inside a layer you don't touch.

The pattern is simple: the more visually separated things were when the image was made, the cleaner they separate coming back out.

Two things worth knowing before you start

Decomposition costs 5 credits per run. Your first runs as a guest are free, which is enough to see whether your image separates well before paying anything.

Export doesn't cost credits — the composite renders locally in your browser. Decompose is the paid step; getting your result out is not.

When this is the wrong tool

If you're generating something new from nothing, use a generator. If you need to remove one scratch from an otherwise perfect photo, a spot-heal tool is faster. Decomposition earns its keep when the image already exists and you need its structure — when the question is "which parts can I change independently" rather than "what should this look like."

The client's winter version took eleven minutes, most of it spent picking the right background. The product never moved a pixel. That's the trade this workflow makes: a little setup in exchange for never re-rolling the parts that already worked.