How I Decompose Any Image Into Recomposable Layers on Melius | Igor Gridel

When you iterate on an AI-generated image by re-prompting, you destroy it.

Ask the model to "move the figure left and remove the truck" and the entire scene shifts: different lighting, different building geometry, different snow. Anything the original seed did not pin gets re-rolled. The bake you liked is gone.

Photographers do not work this way. Concept artists do not work this way. They hold the layers and move them.

I wanted that. So I built it on Melius.

## Clone the canvas

The full pipeline is a public Melius canvas. Free to open, free to test on your own image:

[Open the canvas on Melius](https://app.melius.com/projects/520e4084-2e06-468a-9ea4-6dd7c601b19f/canvas/ea7b354c-65bd-46d5-8ad2-f13ff37816e0)

Drop your input image into the first node, run the canvas, watch every stage produce its output. The exact system prompts on every node are visible inside — read them, change them, fork them.

## The pipeline in one paragraph

One analyzer LLM reads the input image and outputs a JSON blueprint of up to seven candidate layers: background, middle ground, foreground, and up to four subjects. Seven parallel extractor LLMs pull each layer's isolation prompt out of that JSON. Seven NanoBanana Pro nodes regenerate each element on chroma green, preserving original position, scale, and lighting. Background-removal nodes strip the green. A final NanoBanana Pro pass takes all seven transparent layers, accepts new positions for any of them, and renders one unified composition. You can move, scale, drop, or swap any layer before the final pass.

## Why a pipeline and not a prompt

Inpainting and reference conditioning are partial fixes for compositional iteration. The model still has to invent what was behind the thing you moved. Move the truck and the snow underneath gets hallucinated from nothing. Move the figure and the shadow stays where she was standing.

Decomposition pushes the "what was behind it" question into the regeneration stage, where each model call only has to handle one element on chroma green, not a whole scene with occluded geometry.

That requires a real pipeline. A node graph is the right shape for a real pipeline. Melius gave me the canvas that can hold one.

## Read the full breakdown

The full architectural walkthrough lives on Scopeful Pro: every node's job, the design decisions behind the JSON-then-extractor split, why chroma green, why the final regenerative pass, what the canvas is and is not.

[How I Decompose Any Image Into Recomposable Layers on Melius — full breakdown](https://scopeful.org/pro/feed/melius-image-decomposition)

If you fork the canvas and ship a variant, tag me on X ([@IgorGridel](https://x.com/IgorGridel)) — I want to see what other slot taxonomies people land on.