Characterizing class-discriminative signal in visual features: An empirical weight-space study of diffusion model adapters

This paper presents an extended empirical study of weight-space representations for Low-Rank Adapters (LoRAs) in diffusion models, building on our previous ICCS 2025 paper (Duszenko and Bielak, 2025). Our goal is not to introduce a new weight-space learning architecture, but to characterize how much class-discriminative signal can be recovered from simple, interpretable descriptors of LoRA weights, and how this signal changes when text-conditioned layers are removed. We formulate this characterization around a set of empirical working hypotheses concerning signal recoverability, visual-layer retention, descriptor robustness, and the trade-off between interpretability and representational capacity. To evaluate these hypotheses, we compare ten hand-crafted representation methods under a controlled all-layers versus visual-only ablation and use linear classifiers as deliberately low-capacity probes. Ensemble features achieve 68.94% accuracy when using all layers, while visual self-attention layers alone still reach 41.28% accuracy. These results suggest that visual adapter layers retain non-trivial class-related structure under this probing setup, although the observed accuracy should be interpreted as an empirical proxy for separability rather than a direct measure of intrinsic information. Overall, the study provides a clearer empirical characterization of which descriptor families remain robust under layer ablation, with practical implications for efficient adapter analysis and future weight-space modeling.