Reproduce “CHATS: Combining Human-Aligned Optimization and Test-Time Sampling for Text-to-Image Generation” 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
Minghao Fu, Guo-Hua Wang, Liangfu Cao, Qing-Guo Chen, Zhao Xu, Weihua Luo, Kaifu Zhang
The CHATS framework integrates human preference optimization during training with test-time sampling in text-to-image generation, leveraging separate models for preferred and dispreferred distributions. This approach achieves superior image quality and text-image alignment, performing well with only 7,459 high-quality preference pairs from the OIP dataset, demonstrating notable data efficiency.
Key insights
- Unifying human preference optimization during training with test-time sampling, through separate models for preferred and dispreferred distributions, leads to improved text-to-image generation.
- A proxy-prompt-based sampling strategy can efficiently integrate information from both preferred and dispreferred models during inference, enhancing quality and alignment.
- High data efficiency can be achieved by focusing on smaller, high-quality preference datasets when optimizing models by independently modeling preferred and dispreferred distributions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.