Stable-Layers: Fine-Tuning Image Layer Decomposition Models with VLM-Scored Reinforcement Learning

TL;DR

Stable-Layers (Stability AI) takes the pretrained Qwen-Image-Layered decomposer and post-trains it with Flow-GRPO + LoRA, using a VLM as the sole source of reward — no paired ground-truth layered images required. The reward design is the contribution: per-sample CoT scoring on five edit-centric criteria gives narrow within-group variance and starves GRPO of signal, so the paper adds a second grid-based calibration pass in which the VLM re-scores all candidates side-by-side. On the Crello dataset the fine-tuned model produces stronger layer separation, fewer blank/artifact layers, and lower per-layer reconstruction error than the Qwen-Image-Layered base.

Key claims

Pretrained image-layer decomposers can be fine-tuned with reinforcement learning using only VLM feedback — no paired (input image, layered output) supervision — by applying Flow-GRPO with LoRA on top of Qwen-Image-Layered [Abstract].
A VLM judge scoring candidate decompositions in isolation compresses its outputs into a narrow band, leaving GRPO with insufficient within-group variance to learn from [Abstract].
A two-stage reward pipeline — structured per-sample scoring on five edit-centric criteria, followed by a grid-based step where the VLM re-scores all candidates side-by-side — recovers enough variance to drive GRPO updates [Abstract].
Post-training improves layer separation, reduces blank/artifact-heavy layers, and lowers per-layer reconstruction error versus the Qwen-Image-Layered base on the Crello dataset [Abstract].

Method

Stable-Layers starts from a frozen Qwen-Image-Layered checkpoint and attaches a LoRA adapter to the diffusion transformer. For each input image the policy samples multiple candidate layer decompositions; a VLM scores each candidate, and group-relative advantages drive a Flow-GRPO update on the LoRA weights only. The reward is the paper’s central design: a first pass asks the VLM to score each candidate independently along five edit-centric criteria, producing a structured CoT trace whose final scalar is the per-sample reward. Because isolated scoring collapses into a narrow band, a second grid-based calibration pass shows all candidates side-by-side and asks the VLM to re-score them comparatively, restoring within-group variance.

Results

Stable-Layers reports qualitative and quantitative gains over the Qwen-Image-Layered base on the Crello design-template benchmark: stronger separation between layers, fewer near-empty or artifact-dominated layers, and lower per-layer reconstruction error after recomposition [Abstract]. Detailed numbers are not exposed in the abstract; the project page (stability-ai.github.io/stable-layers) is the cited source for visual comparisons.

Why it’s interesting

Two threads on the wiki converge here. First, Stable-Layers is the first filed datapoint of RL post-training applied to the decomposition side of Layered Image/Video Decomposition — Qwen-Image-Layered: Towards Inherent Editability via Layer Decomposition and MRT: Masked Region Transformer for Layered Image Generation and Editing at Scale established the SFT decomposer recipe; Stable-Layers shows the same lineage admits GRPO on top, mirroring what Uniworld-V2: Reinforce Image Editing with Diffusion Negative-aware Finetuning and MLLM Implicit Feedback did for image editing and what Pref-GRPO: Pairwise Preference Reward-based GRPO for Stable Text-to-Image Reinforcement Learning did for T2I. Second, the side-by-side calibration trick is a concrete instance of the reward-design pathology flagged on VLM-as-Evaluator: pointwise VLM judges compress into a narrow band (ERNIE-Image-Aes: Robust Image Aesthetics Scoring with Balanced Category Generalization documents the static-rubric version of the same failure mode). Stable-Layers’ fix — pairwise/grid re-scoring as a calibration head on top of the pointwise CoT — is a candidate solution to one of that page’s open questions about which loss shape best survives downstream GRPO use. The “no paired supervision required” framing also overlaps with Learning an Image Editing Model without Image Editing Pairs — both replace dataset construction with a VLM-judged self-generation loop.