CAPTAIN: Semantic Feature Injection for Memorization Mitigation in Text-to-Image Diffusion Models

TL;DR

CAPTAIN is a training-free method that reduces memorization in text-to-image diffusion models by injecting semantically aligned features into latent regions during denoising, improving privacy without sacrificing image quality.

Contribution

It introduces a novel feature injection framework that mitigates memorization in diffusion models without retraining or degrading prompt alignment.

Findings

Significantly reduces memorization compared to baseline methods

Maintains high prompt fidelity and visual quality

Effective across various datasets and prompts

Abstract

Diffusion models can unintentionally reproduce training examples, raising privacy and copyright concerns as these systems are increasingly deployed at scale. Existing inference-time mitigation methods typically manipulate classifier-free guidance (CFG) or perturb prompt embeddings; however, they often struggle to reduce memorization without compromising alignment with the conditioning prompt. We introduce CAPTAIN, a training-free framework that mitigates memorization by directly modifying latent features during denoising. CAPTAIN first applies frequency-based noise initialization to reduce the tendency to replicate memorized patterns early in the denoising process. It then identifies the optimal denoising timesteps for feature injection and localizes memorized regions. Finally, CAPTAIN injects semantically aligned features from non-memorized reference images into localized latent regions, suppressing memorization while preserving prompt fidelity and visual quality. Our experiments show that CAPTAIN achieves substantial reductions in memorization compared to CFG-based baselines while maintaining strong alignment with the intended prompt.