README.md

metadata

title: Stable Layers
emoji: 🗂️
colorFrom: yellow
colorTo: purple
sdk: gradio
sdk_version: 6.22.0
app_file: app.py
pinned: false
short_description: Split an image into editable RGBA layers
python_version: '3.12'
startup_duration_timeout: 1h
license: other
license_name: stabilityai-community
license_link: https://stability.ai/license
models:
  - StabilityLabs/Stable-Layers
  - Qwen/Qwen-Image-Layered
tags:
  - layer-decomposition
  - rgba
  - image-editing
  - lora

🗂️ Stable Layers

Split any image into a stack of editable RGBA layers — an inpainted background plus one object per layer — with StabilityLabs/Stable-Layers, a LoRA over the Qwen/Qwen-Image-Layered pipeline (20.4B DiT).

What you get

Layers come out back-to-front: layer 0 is the background, repainted behind everything that was lifted off it; each higher layer is one object with real alpha. Stacking them all reproduces the original image.

The viewer on the right is the point of the demo — click a layer to select it, drag it around the canvas, and toggle layers off to watch the composite fall apart and rebuild. Every layer is downloadable as a genuine transparent PNG (single layer or the whole set as a ZIP), ready to drop into an editor.

Inference

Locked to the recipe the authors specify, because deviating from it garbles the decomposition:

Setting	Value
Sampler	Heun (2nd order)
Steps	50
CFG	1.0 (off)
Max dimension	640 px
Layers	4 (2–6 selectable)

The QwenImageLayeredPipeline.__call__ path (Euler + true_cfg_scale=4.0) is the base model's recipe, so this Space reimplements the reference decompose.py denoise loop directly. The adapter is a raw PEFT checkpoint on QwenImageTransformer2DModel whose keys carry no transformer. prefix, so it is applied with PeftModel.from_pretrained(...) + merge_and_unload() rather than load_lora_weights (which silently no-ops on it).

Runs on ZeroGPU at size="xlarge": the transformer (40.9 GB), the Qwen2.5-VL text encoder (16.6 GB) and the RGBA VAE total ~58 GB of bf16 weights, over the 48 GB large slice. Transformer blocks are served from an ahead-of-time-compiled (AOTInductor) package when one is available, with an automatic eager fallback.

Credits

Model by Stability AI, released under the Stability AI Community License. Base pipeline by Qwen. Example images are from the Qwen/Qwen-Image-Layered demo Space (Apache-2.0).