Reproduce “StyleAR: Customizing Multimodal Autoregressive Model for Style-Aligned 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
Yi Wu, Lingting Zhu, Shengju Qian, Lei Liu, Wandi Qiao, Lequan Yu, Bin Li
StyleAR enables multimodal autoregressive models to perform style-aligned text-to-image generation by leveraging a novel binary data curation method and a customized architecture. The system demonstrates a balanced performance in adhering to text prompts (CLIP-T 0.2893) and maintaining style consistency (CLIP-I 0.7456), effectively addressing issues of content leakage observed in other approaches.
Key insights
- Autoregressive models can be effectively trained for style-aligned generation using curated binary image-text data, circumventing the need for complex triplet datasets.
- An optimal ratio of mixed raw and stylized data is crucial for robust style learning, preventing both low consistency and 'overfitting' that leads to content leakage.
- Techniques like Gaussian noise injection in style tokens and SAM-integrated semantic exclusion are effective in decoupling style from content, improving prompt adherence and style fidelity.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.