§ 01
PROBLEM
PROBLEM
What it had to solve.
Cross-attention maps are the go-to interpretability signal for diffusion models — but they are noisy, timestep-dependent, and rarely sharp enough to explain where a concept was painted.
reverseDAAM — diffusion interpretability, token-by-token.
A post-hoc interpretability framework that maps image regions back to prompt tokens in text-to-image diffusion models. Introduces the Word–BBox Alignment Score for comparing how faithfully different models paint what they are told.
Models validated
3
SD, SDXL, IF-XL
Attention layers
16
Cross-attn aggregated
Alignment gain
+22
% vs baseline
Status
Review
IEEE SMC 2026
§ 02
APPROACH
APPROACH
How it works.
reverseDAAM aggregates cross-attention across denoising timesteps with an importance-weighted schedule, then inverts the map to attribute image regions to prompt tokens. The Word–BBox Alignment Score lets us compare models by measuring overlap with ground-truth object boxes.
§ 03
RESULTS
RESULTS
What it delivered.
SDXL aligned objects to prompts 22% better than Stable Diffusion 1.5 on a 400-prompt held-out set — with the gap widening for multi-object compositions.