I created something I’ve always wanted.

An automated image decomposition pipeline that takes any complex scene and breaks it into individually isolated layers, background, middle ground, foreground, and standalone subjects, each on chroma green, ready to be moved, swapped, or remixed. It then reassembles them into a polished new composition.

The workflow links multiple AI models, one analyzes the image and identifies all layers, and another regenerates each element on a green screen.

The result is a complete set of recomposable scene layers from a single input image.

It’s the tool I’ve always wished for when iterating on concept art, remixing scenes, and experimenting with compositions.

Process Steps:

1) Image Analysis: Fed a complex scene into an LLM node with a custom system prompt. The agent analyzed the image and identified 7 potential layers: background (sky), middle ground (structures), foreground (ground), and up to 4 discrete subjects.

2) Prompt Extraction: 7 parallel llm extractor nodes each pull one element's isolation prompt from the analysis JSON. Null subjects output a solid green fallback.

3) Element Isolation: Each prompt + the original image as reference feed into NanoBanana Pro nodes. The model regenerates each element isolated on chroma-green (#00FF00), preserving original position, size, and lighting.

4) Green Screen Removal: Background removal strips the green from each output, producing clean isolated elements.

5) Recomposition: Elements can be repositioned, scaled, or swapped. A final NanoBanana Pro pass regenerates the composed layout into a polished, cohesive scene.