Reproduce “DomainGallery: Few-shot Domain-driven Image Generation by Attribute-centric Finetuning” 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
Yuxuan Duan, Yan Hong, Bo Zhang, Jun Lan, Huijia Zhu, Weiqiang Wang, Jianfu Zhang, Li Niu and 1 more
DomainGallery presents an attribute-centric finetuning method for pretrained text-to-image models, enabling high-fidelity and diverse image generation for specific domains using only a few examples. The approach effectively addresses common challenges such as prior attribute leakage, attribute entanglement, and overfitting, outperforming existing baselines across various few-shot scenarios.
Key insights
- Erasing prior latent attributes from placeholder identifier words before finetuning is crucial for preventing unexpected elements from appearing, particularly when applying domains to new categories.
- Explicitly disentangling domain attributes from categorical attributes during finetuning prevents leakage, allowing for robust application of styles or themes across different content categories.
- A transfer-based similarity consistency loss effectively regularizes few-shot finetuning, maintaining diversity and preventing overfitting by ensuring consistent relative relationships between generated samples.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.