Reproduce “From Sparse to Dense: Multi-View GRPO for Flow Models via Augmented Condition Space” 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
Jiazi Bu, Pengyang Ling, Yujie Zhou, Yibin Wang, Yuhang Zang, Tianyi Wei, Xiaohang Zhan, Jiaqi Wang and 3 more
Researchers at Shanghai Jiao Tong University and collaborators devised Multi-View GRPO (MV-GRPO), a framework for text-to-image flow models that transitions from sparse single-view to dense multi-view reward evaluation. This approach augments the condition space with a Condition Enhancer, achieving improved semantic alignment, visual fidelity, and consistent outperformance against baselines across various reward metrics without costly sample regeneration.
Key insights
- Evaluating generated samples against multiple semantically diverse conditions provides richer and more discriminative advantage estimation signals compared to a single-view approach.
- The Condition Enhancer, particularly the VLM-based variant, can effectively generate high-quality, sample-specific posterior captions that capture nuanced visual semantics.
- The theoretical analysis demonstrates minimal 'probability drift' when optimizing the policy with augmented conditions, validating the efficiency of reusing generated samples without compromising SDE dynamics.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.