Diffusion model that reverse-engineers image layers

What it does Qwen-Image-Layered takes a flat image and tears it into multiple RGBA layers—objects, text, background—each with its own transparency mask. Instead of generating pixels from scratch, it reverse-engineers a layered file you can manipulate independently: recolor one element, delete another, or move it without leaving a smudge. The model runs as a diffusers pipeline built on Qwen2.5-VL and exports to standard PNGs, PSDs, or even PowerPoint files.

The interesting bit The model infers structure behind occlusions and supports variable-depth decomposition: you can ask for three layers or eight, then recursively split any layer further. That means you aren’t stuck with a fixed stack; you can keep drilling down until the granularity suits your edit.

Key highlights

• Decomposes a single image into RGBA layers, enabling independent resize, reposition, and recolor operations
• Supports variable-layer decomposition and recursive splitting of any layer into finer sub-layers
• Text prompts describe the overall scene including occluded objects, but do not explicitly control the semantic content of individual layers
• Released under Apache 2.0 and integrates with diffusers, with vLLM-Omni support for inference
• Ships with a Gradio interface that exports decomposed layers to PSD, PPTX, and ZIP formats

Caveats

• Text prompts describe the full scene, not individual layers; you cannot explicitly dictate what each layer contains
• The released weights are tuned for image-to-layer decomposition, so text-to-multi-RGBA generation from scratch is limited

Verdict Worth a look if you build image editing tools, design automation pipelines, or just want to stop manually cutting out objects. Skip it if you need a text-to-layer generator from scratch—the model is trained to decompose existing images, not hallucinate layered scenes from a blank canvas.