ASemConsist: Training-Free Identity Consistency in Text-to-Image Generation

This paper addresses the critical challenge of maintaining character identity consistency across multiple images generated from text prompts using diffusion models. It proposes a novel framework, ASemConsist, that achieves this without requiring any training, a significant advantage. The core contributions include selective text embedding modification, repurposing padding embeddings for semantic control, and an adaptive feature-sharing strategy. The introduction of the Consistency Quality Score (CQS) provides a unified metric for evaluating performance, addressing the trade-off between identity preservation and prompt alignment. The paper's focus on a training-free approach and the development of a new evaluation metric are particularly noteworthy.