Denoising Models Develop Human-Like Perceptual Illusion Representations Across Architectures

Project website →

Contents

Directory	Description
data/	Full effect-size tables, per-channel rankings, ablation results, psychophysics, permutation p-values, shift fractions
code/	Probing pipeline, ablation, injection, psychophysics, statistics utilities
docs/	Source for the project website (README)

---

data/

effect_sizes/

Full layer × timestep tables with Cohen's d, Hedge's g, and Glass's Δ for every model. All metrics are signed; use abs() for ranking. t-test p-values are included for completeness only — effect size is the primary result criterion throughout the paper.

File	Model	Layers	Timesteps
church_ddpm_full.csv	google/ddpm-ema-church-256	41	6
bedroom_ddpm_full.csv	google/ddpm-ema-bedroom-256	41	6
celebahq_ddpm_full.csv	google/ddpm-ema-celebahq-256	41	6
ldm_celebahq_full.csv	LDM (CelebA-HQ)	41	6
sd15_full.csv	runwayml/stable-diffusion-v1-5 (trained + 3 random seeds)	32	5
dit_xl2_full.csv	DiT-XL/2	28	6
resnet50_full.csv	ResNet-50 (ImageNet)	17	—
vgg19_full.csv	VGG-19 (ImageNet)	19	—
vit_b16_full.csv	ViT-B/16 (ImageNet)	12	—
vit_l16_full.csv	ViT-L/16 (ImageNet)	24	—

Columns (DDPM models): layer, timestep, category, n_images, cohens_d, hedges_g, glass_delta, p_value, p_fdr, significant

Columns (CNNs/ViTs): image, layer, timestep, mean_ill, mean_ctrl, delta, model

per_channel/

Per-channel effect sizes at t=50 for the three key layers (N=72 color images).

Columns: layer, channel, cohens_d, hedges_g, glass_delta, p_value, n_images, p_fdr, significant

File	Layer	Channels
mid_attn_0_per_channel.csv	mid_attn_0	512
mid_resnet_0_per_channel.csv	mid_resnet_0	512
down_5_resnet_1_per_channel.csv	down_5_resnet_1	512

Channel 311 is the top channel by |d| across all three layers (see paper Section 3.4).

ablation/

File	Description	Paper reference
threshold_sweep.csv	% Δ reduction vs d-threshold (0.2–0.8) for pos_d / neg_d groups	Appendix W
cv_results.csv	5-fold + LOO cross-validation of channel selection stability	Appendix AH
random_null_500.csv	500 random-channel ablation samples (null distribution)	Appendix
mse_phantom_test.csv	MSE/MAE phantom test — ablated vs random channels on generated images	Appendix AP

psychophysics/

File	Description
flodog_per_layer_rho.csv	Spearman ρ between FLODOG brightness predictions and per-layer activations (color illusions)
dose_response.csv	Mean activation delta vs illusion amplitude for Ebbinghaus + Ponzo at 7 strength levels

permutation/

File	Description
permutation_pvalues_246.csv	FDR-corrected permutation p-values for all 246 (layer × timestep) combinations

shift_fractions/

File	Description
all_models_shift_frac.csv	Phantom injection shift fractions for church-DDPM, DiT-XL/2, and LDM-CelebA-HQ

---

code/

See code/README.md for full reproduction instructions.

probing/
  hooks.py            activation capture via forward hooks
  effect_sizes.py     Cohen's d, Hedge's g, Glass's Δ
  stimuli.py          GVIL + programmatic illusion loaders
  run_probe.py        main probing entry point

ablation/
  channel_ablation.py
  cross_validation.py
  random_null.py

injection/
  ddim_injection.py
  single_step_injection.py
  within_manifold.py

psychophysics/
  flodog.py
  dose_response.py

stats/
  bootstrap.py
  permutation.py

---

Data schemas

Effect size columns

• cohens_d — paired Cohen's d = mean(Δ) / std(Δ)
• hedges_g — bias-corrected; preferred for N < 50 (g ≈ 0.978 × d at N=35)
• glass_delta — normalized by control-region std; independent of illusion-region variance

Region definitions

• illusion_a / bbox_match — humanlike illusion region (appears perceptually different)
• illusion_b / bbox_mismatch — control region (matched location, no illusory context)
• Δ = mean activation in illusion_a − mean activation in illusion_b, per image, then paired

Models not included

Model weights are not shipped. All models load automatically from HuggingFace:

• google/ddpm-ema-church-256, google/ddpm-ema-bedroom-256, google/ddpm-ema-celebahq-256
• runwayml/stable-diffusion-v1-5
• DiT-XL/2: facebook/DiT-XL-2-256