Browse Teach Curate

Nested Attention for Subject Personalization in Text-to-Image Models

A personalization mechanism for text-to-image diffusion that ties a subject to a single text token in the standard cross-attention layer but replaces the fixed attention value with query-dependent subject values produced by an inner (nested) attention layer, selecting relevant subject features per image region to preserve identity while maintaining the model's prior and enabling multiple subjects in one image.

B9 Text

Nested Attention: Semantic-aware Attention Values for Concept Personalization Or Patashnik†,§ Rinon

This paper addresses subject personalization of text-to-image diffusion models: generating images of a specific subject (a pet, a person) across varied scenes and styles. The core tension is between …