Reproduce “Stable-Layers: Fine-Tuning Image Layer Decomposition Models with VLM-Scored Reinforcement Learning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Ciara Rowles, Reshinth Adithyan, Nikhil Pinnaparaju, Vikram Voleti, Mark Boss
Stable-Layers refines an existing image-layer model by using a vision-language model to score candidate RGBA decompositions and compare them side by side, then updating the model with group-relative rewards. On 480 held-out LAION-Aesthetics images, blank or glazed layers fall from about 1.65 to about 0.4 per decomposition during training, while automated evaluations do not establish improved human editing outcomes.
Key insights
- High-quality image layer decomposition models can be effectively fine-tuned using VLM-as-judge reinforcement learning alone, circumventing the need for expensive and ambiguous paired ground-truth datasets.
- A two-phase VLM reward system, incorporating explicit relative judgments via grid calibration, is crucial for extracting a discriminative learning signal for complex, multi-dimensional quality criteria like decomposition, even when absolute scores for candidates are similar.
- Over-specifying an idealized decomposition in text prompts can hinder the policy model's learning, suggesting that simpler, basic prompts might be more effective for guiding VLM feedback in certain generative tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.