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

Key points

• We therefore reformulate diffusion data attribution as attributing factor-induced internal response trajectories.1
• In this paper, we propose a novel Concept Attribution method through Dynamic Trajectories(CADT).1
• Specifically, we construct matched counterfactual pairs at identical noisy states to isolate factor-specific representation displacements, and model their directional and magnitude evolution across denoising as dynamic attribution signatures.1
• Experiments on multiple public datasets show consistent improvements over existing diffusion attribution baselines across hierarchical, compositional, and style attribution.1

Sources (1)

• [1]Attributing HOW, Not Just WHICH: Counterfactual Response Trajectories for Diffusion Models
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 8, 04:40 AM

  Diffusion models have achieved remarkable success in image generation, yet tracing their outputs to individual training examples remains challenging.

  We therefore reformulate diffusion data attribution as attributing factor-induced internal response trajectories.

Extractive summary: sentences quoted from the sources.