Reproduce “TextSphere: Efficient Text-to-Image Generation via Shared Hyperspherical Alignment and Uniformity” 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
Yixia Zhang, Girish A. Koushik, Xilu Wang, Wenwu Wang
TextSphere adapts Sphere Encoder for text-to-image generation by conditioning on captions and explicitly aligning text and spherical image representations. It uses a frozen pretrained CLIP text encoder and an alignment objective alongside reconstruction losses. On CUB-200-2011, it reports gFID 8.03 at four steps, versus 10.21 for Improved VQ-Diffusion at 100 steps; this does not establish lower runtime or broader generalization.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.