TextSphere: Efficient Text-to-Image Generation via Shared Hyperspherical Alignment and Uniformity
TextSphere reports better generation scores than the compared models trained from scratch, using just four or six sampling steps.
The results cover CUB-200-2011 at four steps and MS-COCO-2014 at six, measured by generation FID, where lower is better. The framework adds pretrained CLIP text embeddings to Sphere Encoder and introduces an objective that aligns text and image representations on a shared hypersphere. Across three datasets, it improves generation FID over the text-conditioned baseline while keeping reconstruction FID below 1.0. The authors also find that better reconstruction fidelity does not necessarily mean better generated images.
ArXiv · AI/CL/LG's note
The results cover CUB-200-2011 at four steps and MS-COCO-2014 at six, measured by generation FID, where lower is better. The framework adds pretrained CLIP text embeddings to Sphere Encoder and introduces an objective that aligns text and image representations on a shared hypersphere. Across three datasets, it improves generation FID over the text-conditioned baseline while keeping reconstruction FID below 1.0. The authors also find that better reconstruction fidelity does not necessarily mean better generated images.
ArXiv · AI/CL/LG's note
score 4