DiffEx

DiffEx (CVPR 2025) – Official Repository

This repository contains the code for DiffEx, a diffusion-based framework for explaining image classifiers using semantic image edits.

pip install torch torchvision transformers diffusers pillow

Use this only if you want to automatically extract high-level semantic categories from images.

python parent_semantic_ext.py

Converts flat categories from first step output into nested, fine-grained semantic keywords.

python expand_categories.py \
  --in_json data_parent_semantic.json \
  --out_json data_nested.json

Output: nested taxonomy JSON (required by DiffEx).

Main script. Applies diffusion-based edits and measures classifier sensitivity.

python diffex_run.py \
  --taxonomy_json data_nested.json \
  --image_path 0004.png \
  --out_json diffex_output.json \
  --clip_labels "male" "female" \
  --clip_target_index 1