Reproduce “Di3PO - Diptych Diffusion DPO for Targeted Improvements in 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
Sanjana Reddy, Ishaan Malhi, Sally Ma, Praneet Dutta
Di3PO trains image generators to render text more accurately by constructing preference pairs with correctly spelled and misspelled text against similar backgrounds. It generates panels from one prompt, filters them, and uses the correct panel as the preferred example. On 2,000 prompts, SDXL 1.0’s word error rate was 0.6456 versus 0.7213 for SFT; results test text rendering only.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.