Inject Where It Matters: Training-Free Spatially-Adaptive Identity Preservation for Text-to-Image Personalization

TL;DR

SpatialID is a training-free method that adaptively injects identity features into specific regions during text-to-image generation, improving identity preservation and background consistency without fine-tuning.

Contribution

It introduces a spatially-adaptive identity modulation framework and a dynamic scheduling strategy, addressing background contamination issues in personalization.

Findings

Achieves state-of-the-art text adherence and visual consistency.

Significantly reduces background contamination in generated images.

Maintains robust identity preservation without fine-tuning.

Abstract

Personalized text-to-image generation aims to integrate specific identities into arbitrary contexts. However, existing tuning-free methods typically employ Spatially Uniform Visual Injection, causing identity features to contaminate non-facial regions (e.g., backgrounds and lighting) and degrading text adherence. To address this without expensive fine-tuning, we propose SpatialID, a training-free spatially-adaptive identity modulation framework. SpatialID fundamentally decouples identity injection into face-relevant and context-free regions using a Spatial Mask Extractor derived from cross-attention responses. Furthermore, we introduce a Temporal-Spatial Scheduling strategy that dynamically adjusts spatial constraints - transitioning from Gaussian priors to attention-based masks and adaptive relaxation - to align with the diffusion generation dynamics. Extensive experiments on IBench demonstrate that SpatialID achieves state-of-the-art performance in text adherence (CLIP-T: 0.281), visual consistency (CLIP-I: 0.827), and image quality (IQ: 0.523), significantly eliminating background contamination while maintaining robust identity preservation.