Researchers have developed SEAL, a new method for personalizing stickers in text-to-image generation using a single reference image. SEAL addresses issues like visual entanglement and structural rigidity that arise with existing test-time fine-tuning methods. The approach integrates as a plug-and-play module into diffusion models and utilizes a Semantic-guided Spatial Attention Loss, a Split-merge Token Strategy, and Structure-aware Layer Restriction. To facilitate evaluation, a large-scale dataset called StickerBench has been created with structured tags for attribute-level control. AI
Summary written by gemini-2.5-flash-lite from 2 sources. How we write summaries →
IMPACT Introduces a novel technique for personalized image generation, potentially improving control and reducing artifacts in sticker creation.
RANK_REASON This is a research paper describing a new method and dataset for image generation.