Stable-Layers: Fine-Tuning an Image Layer Decomposition Model via Reinforcement Learning Using VLM Scores

Stable-Layers is a reinforcement learning framework that fine-tunes a pretrained layer decomposition model without paired supervision. It is implemented based on Qwen-Image-Layered, applying Flow-GRPO and LoRA adaptation techniques.

The framework samples multiple candidate decompositions per image, scores them with a VLM, and optimizes the policy through group-relative advantages. However, when the VLM evaluates individual samples independently, the score variance narrows, resulting in insufficient variability for GRPO training.

To address this, a two-stage evaluation pipeline is proposed:

• Structured individual sample scoring based on 5 edit-focused criteria
• Grid-based calibration where the VLM re-scores all candidates by comparing them side by side

Experimental results on the Crello dataset show enhanced layer separation performance compared to the existing model, along with fewer empty layers or artifacts and lower per-layer reconstruction error.