Reproduce “Diversify, Don't Fine-Tune: Scaling Up Visual Recognition Training with Synthetic Images” 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
Zhuoran Yu, Chenchen Zhu, Sean Culatana, Raghuraman Krishnamoorthi, Fanyi Xiao, Yong Jae Lee
The authors present a recipe for using generated images to improve ImageNet classification. They keep an off-the-shelf diffusion model fixed, diversify prompts, and balance real and synthetic images during training. ResNet-50 reaches 79.27% top-1 accuracy, versus 76.74% with real-only training. Gains continue through six times the training-set size, but the cross-study comparison does not isolate fine-tuning as the cause of declines.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.