Attributing HOW, Not Just WHICH: Counterfactual Response Trajectories for Diffusion Models

arXiv cs.GR details a new diffusion-attribution approach that shifts the question from which training examples influenced an output to how that influence unfolds during denoising. The paper introduces Concept Attribution through Dynamic Trajectories, or CADT, for tracing factor-specific internal response trajectories in image generators.

The core idea is to compare matched counterfactual pairs at the same noisy state, then track directional and magnitude changes in representation across denoising steps. Those stage-wise features are integrated into trajectory descriptors for both the query image and the training set, giving the method a richer signal than scalar attribution scores.

For developers working on generative tools, this matters because provenance, dataset debugging, and style tracing all get harder as diffusion models become more capable and more opaque. A trajectory-based view could help teams identify which examples are driving unwanted stylistic bleed, compositional quirks, or hierarchy errors in outputs.

The paper says experiments on multiple public datasets show consistent gains over existing diffusion attribution baselines across hierarchical, compositional, and style attribution. The exact production implications are still research-stage, but the direction is clear: attribution for generative models is moving toward temporal, factor-aware explanations rather than one-number similarity scores.