Stable-Layers: Fine-Tuning Image Layer Decomposition Models with VLM-Scored Reinforcement Learning
Researchers have developed Stable-Layers, a novel reinforcement learning framework designed to improve image layer decomposition models. This system bypasses the need for paired training data by utilizing feedback from a vision-language model (VLM). By employing Flow-GRPO and LoRA adaptation, Stable-Layers optimizes policy training and has demonstrated enhanced layer separation and reduced reconstruction errors on the Crello dataset compared to its base model. AI
IMPACT Introduces a method for improving image decomposition models without paired data, potentially reducing data annotation costs.