dinac3 Reconstruction Results

Side-by-side reconstructions, RGB deltas, and latent PCA for dinac3_96 on the fixed 39-image comparison set.

Metricdinac3_96
Mean PSNR (dB)32.05
Median PSNR (dB)30.82
Min PSNR (dB)22.57
Max PSNR (dB)44.31
Per-image PSNR (dB), 39 images
Imagedinac3_96
p640x1536:9462327.81
p640x1536:9462430.29
p640x1536:9462539.08
p640x1536:9462638.02
p640x1536:9462738.59
p640x1536:9462840.46
p960x1024:21626439.60
p960x1024:21626535.18
p960x1024:21626644.31
p960x1024:21626735.94
p960x1024:21626825.71
p960x1024:21626938.52
p960x1024:21627033.49
p960x1024:21627137.64
p704x1472:9469927.26
p704x1472:9470025.09
p704x1472:9470137.33
p704x1472:9470230.36
p704x1472:9470331.44
p704x1472:9470428.55
p704x1472:9470540.42
p704x1472:9470639.18
r256_p1344x704:1557726.13
r256_p1344x704:1557825.84
r256_p1344x704:1557927.43
r256_p1344x704:1558032.11
r256_p1344x704:1558130.06
r256_p1344x704:1558226.71
r256_p1344x704:1558329.72
r256_p1344x704:1558426.24
r256_p896x1152:14413128.38
r256_p896x1152:14413230.90
r256_p896x1152:14413333.89
r256_p896x1152:14413429.84
r256_p896x1152:14413525.10
r256_p896x1152:14413630.82
r256_p896x1152:14413728.14
r256_p896x1152:14413822.57
exports/dico_diffae_hf/VAE_accuracy_test_image.png31.83
Inference settings
SettingValue
Decodeone deterministic pass (no sampler, no seed)
Precisionbfloat16 (stored weights, BF16 autocast); FP32 latents/output
Compiletorch.compile encoder and decoder (package default)
InputFP32 RGB in [-1, 1], center-cropped to multiples of 16
PSNRreconstruction clamped to [-1, 1] vs FP32 input, peak-to-peak 2

benchmark.json ยท _results_appendix.md

Image width:
Reconstruction
RGB Difference
Latent PCA

p640x1536:94623

Original
dinac3_96

PSNR 27.81 dB

p640x1536:94624

Original
dinac3_96

PSNR 30.29 dB

p640x1536:94625

Original
dinac3_96

PSNR 39.08 dB

p640x1536:94626

Original
dinac3_96

PSNR 38.02 dB

p640x1536:94627

Original
dinac3_96

PSNR 38.59 dB

p640x1536:94628

Original
dinac3_96

PSNR 40.46 dB

p960x1024:216264

Original
dinac3_96

PSNR 39.60 dB

p960x1024:216265

Original
dinac3_96

PSNR 35.18 dB

p960x1024:216266

Original
dinac3_96

PSNR 44.31 dB

p960x1024:216267

Original
dinac3_96

PSNR 35.94 dB

p960x1024:216268

Original
dinac3_96

PSNR 25.71 dB

p960x1024:216269

Original
dinac3_96

PSNR 38.52 dB

p960x1024:216270

Original
dinac3_96

PSNR 33.49 dB

p960x1024:216271

Original
dinac3_96

PSNR 37.64 dB

p704x1472:94699

Original
dinac3_96

PSNR 27.26 dB

p704x1472:94700

Original
dinac3_96

PSNR 25.09 dB

p704x1472:94701

Original
dinac3_96

PSNR 37.33 dB

p704x1472:94702

Original
dinac3_96

PSNR 30.36 dB

p704x1472:94703

Original
dinac3_96

PSNR 31.44 dB

p704x1472:94704

Original
dinac3_96

PSNR 28.55 dB

p704x1472:94705

Original
dinac3_96

PSNR 40.42 dB

p704x1472:94706

Original
dinac3_96

PSNR 39.18 dB

r256_p1344x704:15577

Original
dinac3_96

PSNR 26.13 dB

r256_p1344x704:15578

Original
dinac3_96

PSNR 25.84 dB

r256_p1344x704:15579

Original
dinac3_96

PSNR 27.43 dB

r256_p1344x704:15580

Original
dinac3_96

PSNR 32.11 dB

r256_p1344x704:15581

Original
dinac3_96

PSNR 30.06 dB

r256_p1344x704:15582

Original
dinac3_96

PSNR 26.71 dB

r256_p1344x704:15583

Original
dinac3_96

PSNR 29.72 dB

r256_p1344x704:15584

Original
dinac3_96

PSNR 26.24 dB

r256_p896x1152:144131

Original
dinac3_96

PSNR 28.38 dB

r256_p896x1152:144132

Original
dinac3_96

PSNR 30.90 dB

r256_p896x1152:144133

Original
dinac3_96

PSNR 33.89 dB

r256_p896x1152:144134

Original
dinac3_96

PSNR 29.84 dB

r256_p896x1152:144135

Original
dinac3_96

PSNR 25.10 dB

r256_p896x1152:144136

Original
dinac3_96

PSNR 30.82 dB

r256_p896x1152:144137

Original
dinac3_96

PSNR 28.14 dB

r256_p896x1152:144138

Original
dinac3_96

PSNR 22.57 dB

exports/dico_diffae_hf/VAE_accuracy_test_image.png

Original
dinac3_96

PSNR 31.83 dB

Absolute RGB error with shared 99.5th-percentile normalization and sqrt scaling. Brighter = more error.

p640x1536:94623

Original
dinac3_96

PSNR 27.81 dB

p640x1536:94624

Original
dinac3_96

PSNR 30.29 dB

p640x1536:94625

Original
dinac3_96

PSNR 39.08 dB

p640x1536:94626

Original
dinac3_96

PSNR 38.02 dB

p640x1536:94627

Original
dinac3_96

PSNR 38.59 dB

p640x1536:94628

Original
dinac3_96

PSNR 40.46 dB

p960x1024:216264

Original
dinac3_96

PSNR 39.60 dB

p960x1024:216265

Original
dinac3_96

PSNR 35.18 dB

p960x1024:216266

Original
dinac3_96

PSNR 44.31 dB

p960x1024:216267

Original
dinac3_96

PSNR 35.94 dB

p960x1024:216268

Original
dinac3_96

PSNR 25.71 dB

p960x1024:216269

Original
dinac3_96

PSNR 38.52 dB

p960x1024:216270

Original
dinac3_96

PSNR 33.49 dB

p960x1024:216271

Original
dinac3_96

PSNR 37.64 dB

p704x1472:94699

Original
dinac3_96

PSNR 27.26 dB

p704x1472:94700

Original
dinac3_96

PSNR 25.09 dB

p704x1472:94701

Original
dinac3_96

PSNR 37.33 dB

p704x1472:94702

Original
dinac3_96

PSNR 30.36 dB

p704x1472:94703

Original
dinac3_96

PSNR 31.44 dB

p704x1472:94704

Original
dinac3_96

PSNR 28.55 dB

p704x1472:94705

Original
dinac3_96

PSNR 40.42 dB

p704x1472:94706

Original
dinac3_96

PSNR 39.18 dB

r256_p1344x704:15577

Original
dinac3_96

PSNR 26.13 dB

r256_p1344x704:15578

Original
dinac3_96

PSNR 25.84 dB

r256_p1344x704:15579

Original
dinac3_96

PSNR 27.43 dB

r256_p1344x704:15580

Original
dinac3_96

PSNR 32.11 dB

r256_p1344x704:15581

Original
dinac3_96

PSNR 30.06 dB

r256_p1344x704:15582

Original
dinac3_96

PSNR 26.71 dB

r256_p1344x704:15583

Original
dinac3_96

PSNR 29.72 dB

r256_p1344x704:15584

Original
dinac3_96

PSNR 26.24 dB

r256_p896x1152:144131

Original
dinac3_96

PSNR 28.38 dB

r256_p896x1152:144132

Original
dinac3_96

PSNR 30.90 dB

r256_p896x1152:144133

Original
dinac3_96

PSNR 33.89 dB

r256_p896x1152:144134

Original
dinac3_96

PSNR 29.84 dB

r256_p896x1152:144135

Original
dinac3_96

PSNR 25.10 dB

r256_p896x1152:144136

Original
dinac3_96

PSNR 30.82 dB

r256_p896x1152:144137

Original
dinac3_96

PSNR 28.14 dB

r256_p896x1152:144138

Original
dinac3_96

PSNR 22.57 dB

exports/dico_diffae_hf/VAE_accuracy_test_image.png

Original
dinac3_96

PSNR 31.83 dB

Top 3 PCA components of each model's whitened encoder latent mapped to RGB, per image. Colors of each block are aligned to the first column via spatial correlation matching.

p640x1536:94623

Original
dinac3_96

p640x1536:94624

Original
dinac3_96

p640x1536:94625

Original
dinac3_96

p640x1536:94626

Original
dinac3_96

p640x1536:94627

Original
dinac3_96

p640x1536:94628

Original
dinac3_96

p960x1024:216264

Original
dinac3_96

p960x1024:216265

Original
dinac3_96

p960x1024:216266

Original
dinac3_96

p960x1024:216267

Original
dinac3_96

p960x1024:216268

Original
dinac3_96

p960x1024:216269

Original
dinac3_96

p960x1024:216270

Original
dinac3_96

p960x1024:216271

Original
dinac3_96

p704x1472:94699

Original
dinac3_96

p704x1472:94700

Original
dinac3_96

p704x1472:94701

Original
dinac3_96

p704x1472:94702

Original
dinac3_96

p704x1472:94703

Original
dinac3_96

p704x1472:94704

Original
dinac3_96

p704x1472:94705

Original
dinac3_96

p704x1472:94706

Original
dinac3_96

r256_p1344x704:15577

Original
dinac3_96

r256_p1344x704:15578

Original
dinac3_96

r256_p1344x704:15579

Original
dinac3_96

r256_p1344x704:15580

Original
dinac3_96

r256_p1344x704:15581

Original
dinac3_96

r256_p1344x704:15582

Original
dinac3_96

r256_p1344x704:15583

Original
dinac3_96

r256_p1344x704:15584

Original
dinac3_96

r256_p896x1152:144131

Original
dinac3_96

r256_p896x1152:144132

Original
dinac3_96

r256_p896x1152:144133

Original
dinac3_96

r256_p896x1152:144134

Original
dinac3_96

r256_p896x1152:144135

Original
dinac3_96

r256_p896x1152:144136

Original
dinac3_96

r256_p896x1152:144137

Original
dinac3_96

r256_p896x1152:144138

Original
dinac3_96

exports/dico_diffae_hf/VAE_accuracy_test_image.png

Original
dinac3_96