DivRL: Disentangled Self-Similarity Rewards for Diverse Subject-Driven Generation

2026-06-22

AI Summary

This work addresses the “identity-diversity paradox” in topic-driven image generation, where maintaining identity consistency often comes at the expense of structural diversity. To resolve this trade-off, the authors propose DivRL, a novel framework that decouples visual features and formulates identity preservation as a feasibility constraint. DivRL employs an exploration-suppression strategy to jointly optimize diversity and consistency, introducing two key innovations: negative self-similarity metric (nSSM) and visual-semantic matching (VSM). These components, combined with a gated optimization mechanism and a squared hinge loss, enable significant gains in structural diversity without compromising identity fidelity. Experimental results demonstrate that DivRL effectively overcomes the performance bottleneck inherent in existing approaches to the identity-diversity trade-off.

Abstract

Subject-driven image generation faces an "Identity-Diversity Paradox", where strong identity preservation often leads to rigid and low-diversity outputs. We propose a post-training framework called DivRL that jointly optimizes identity consistency and structural diversity simultaneously by leveraging disentangled visual features from a robust similarity model. Specifically, we introduce a Negative Self-Similarity Measure (nSSM) to quantify structural diversity, and Visual Semantic Matching (VSM) to evaluate identity consistency. We propose an "Explore-and-Suppress" strategy that treats VSM as a gated constraint: the model freely explores structurally diverse configurations, and only samples that violate the identity threshold are penalized via a quadratic hinge loss. This converts identity preservation from a competing objective into a feasibility constraint, allowing nSSM and VSM to improve jointly. Experiments demonstrate that our method effectively pushes the model to generate both consistent and diverse images and improves structural diversity while maintaining comparable identity consistency through a gated optimization formulation.

Problem

• Subject-driven generation
• Identity-Diversity Paradox
• Structural diversity
• Identity consistency

Innovation

• Disentangled Self-Similarity
• Subject-Driven Generation
• Identity-Diversity Paradox
• Negative Self-Similarity Measure
• Visual Semantic Matching