Stability AI releases Stable-Layers, a reinforcement learning framework that trains image layer decomposition models without paired supervision data using Flow-GRPO and VLM feedback.

Stability AI has introduced Stable-Layers, a reinforcement learning framework designed to train image layer decomposition models without requiring paired training datasets. Traditionally, splitting a flat image into editable, multi-layer components required intensive human annotation. Stable-Layers bypasses this by adapting Group Relative Policy Optimization (GRPO) for flow-matching models (Flow-GRPO) to optimize image decomposition. The training is guided by a Vision-Language Model (VLM) serving as a judge, using a structured scoring and grid-based calibration pipeline to provide high-quality reward signals. This approach significantly reduces color bleed and blank layer artifacts, producing cleaner semantic separation.