Reproduce “Accelerating Inference of Masked Image Generators via 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
Pranav Subbaraman, Shufan Li, Siyan Zhao, Aditya Grover
Speed-RL is introduced as a reinforcement learning framework designed to accelerate masked image generators (MGMs). This method achieves up to a 3x inference speedup for base MGMs like Meissonic, while enhancing image quality through a novel combined reward and an adapted Group-Ratio Policy Optimization algorithm.
Key insights
- Directly optimizing for both image quality and inference speed through reinforcement learning provides a more effective acceleration strategy for MGMs compared to distribution-matching approaches.
- The stability of RL training for discrete masked generative models is significantly improved by using a low-variance KL estimator that leverages full token distributions rather than just sampled tokens.
- Filtering low-quality samples and frequently updating the policy with fresh samples are crucial for maintaining training stability and maximizing performance gains in this RL-driven acceleration framework.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.