Stable Layers

Stable-Layers is a research framework from Stability AI for decomposing an image into separate transparent layers. It starts from the pretrained Qwen-Image-Layered model and fine-tunes it with Flow-GRPO and LoRA adaptation, using feedback from a vision-language model (VLM) instead of paired ground-truth supervision. Training uses only unlabelled images.

For each image, multiple candidate decompositions are sampled and scored by a VLM, and the policy is updated from group-relative advantages. To avoid the VLM compressing its scores into a narrow band, the authors use a two-stage evaluation: structured per-sample scoring on five edit-centric criteria, followed by a grid-based calibration step in which the VLM re-scores all candidates side by side.

According to the project page, compared with the base model it gives stronger layer separation, fewer blank or artifact-heavy layers, cleaner alpha masks, and lower per-layer L1 reconstruction error, with the largest gains on the first layer. The base model often leaves Layer 0 degenerate and duplicates the composite across foreground slots. At the time of the provided material, only a project page and an arXiv paper are referenced; no code, weights or hosted demo links were provided.