byronliang8/diffusion-geometry

This research tool evaluates diffusion-model instability by comparing prompt-conditioned generations, estimating local geometry from finite-difference perturbations, and saving diagnostic maps or per-sample metrics. The main workflow is documented for paired in-distribution and out-of-distribution prompts; the sampled code shows SD3.5 prompt encoding, sampling, low-dimensional Jacobian estimation, and metric computation. Heatmap rendering and result writing are documented, though their complete implementations were not sampled. Both scripts support distributed execution; the graph focuses on product behavior, not execution infrastructure.