DriftScope: Measuring The Hidden Effects of Diffusion Model Adaptation

Adapting pre-trained text-to-image diffusion models is a standard practice for customizing visual concepts or removing unwanted content. However, these weight-level modifications are typically evaluated only on their intended effects—whether the new concept is learned or the target concept is erased. This paper demonstrates that this narrow evaluation framework masks significant collateral damage. Even when aggregate metrics like FID and KID remain stable, adapted models often suffer from severe, localized degradation of unrelated concepts, effectively creating "conceptual blindspots" that go unnoticed until deployment.

The authors provide empirical evidence of this phenomenon using two complementary approaches. First, they employ sparse autoencoders (SAEs) to analyze shifts in the generative distribution, revealing that adaptation induces heavy-tailed changes in concept prevalence. Second, they use zero-shot classification via the Gaussian Diffusion Classifier (GDC) to confirm that these shifts translate into measurable losses in semantic capability. They find that while some methods cause catastrophic, easily detectable failure, others—such as SPM or DreamBooth—induce insidious damage where mean performance appears stable, but worst-case per-class accuracy drops significantly. This indicates that standard aggregate metrics are structurally blind to concentrated, concept-specific erosion.

To address this, the authors introduce DriftScope, a prompt-level diagnostic tool designed to audit models before they are deployed. DriftScope takes two model checkpoints—a base model and an adapted version—and optimizes a soft prompt to maximize the divergence between their cross-attention maps. By identifying which tokens produce the most divergent attention patterns, the tool generates a ranked "blacklist" of concepts that have been most affected by the adaptation. Because DriftScope is differentiable and operates on the prompt level, it provides an interpretable, concept-specific report without requiring access to internal model weights or real training data.