iRDiffAE — Reconstruction Results

Comparison of iRDiffAE model variants and Flux VAE references.

Image width:
Reconstruction
RGB Difference
Latent PCA

p640x1536:94623

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p640x1536:94624

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p640x1536:94625

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p640x1536:94626

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p640x1536:94627

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p640x1536:94628

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p960x1024:216264

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p960x1024:216265

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p960x1024:216266

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p960x1024:216267

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p960x1024:216268

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p960x1024:216269

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p960x1024:216270

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p960x1024:216271

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p704x1472:94699

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p704x1472:94700

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p704x1472:94701

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p704x1472:94702

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p704x1472:94703

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p704x1472:94704

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p704x1472:94705

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p704x1472:94706

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

r256_p1344x704:15577

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

r256_p1344x704:15578

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

r256_p1344x704:15579

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

r256_p1344x704:15580

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

r256_p1344x704:15581

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

r256_p1344x704:15582

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

r256_p1344x704:15583

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

r256_p1344x704:15584

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

r256_p896x1152:144131

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

r256_p896x1152:144132

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

r256_p896x1152:144133

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

r256_p896x1152:144134

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

r256_p896x1152:144135

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

r256_p896x1152:144136

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

r256_p896x1152:144137

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

r256_p896x1152:144138

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

VAE_accuracy_test_image

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

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

p640x1536:94623

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p640x1536:94624

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p640x1536:94625

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p640x1536:94626

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p640x1536:94627

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p640x1536:94628

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p960x1024:216264

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p960x1024:216265

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p960x1024:216266

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p960x1024:216267

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p960x1024:216268

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p960x1024:216269

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p960x1024:216270

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p960x1024:216271

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p704x1472:94699

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p704x1472:94700

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p704x1472:94701

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p704x1472:94702

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p704x1472:94703

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p704x1472:94704

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p704x1472:94705

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p704x1472:94706

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

r256_p1344x704:15577

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

r256_p1344x704:15578

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

r256_p1344x704:15579

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

r256_p1344x704:15580

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

r256_p1344x704:15581

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

r256_p1344x704:15582

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

r256_p1344x704:15583

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

r256_p1344x704:15584

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

r256_p896x1152:144131

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

r256_p896x1152:144132

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

r256_p896x1152:144133

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

r256_p896x1152:144134

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

r256_p896x1152:144135

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

r256_p896x1152:144136

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

r256_p896x1152:144137

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

r256_p896x1152:144138

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

VAE_accuracy_test_image

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

Top 3 PCA components of encoder latents mapped to RGB. Colors are aligned across models via spatial correlation matching. Reruns (same encoder) are omitted.

p640x1536:94623

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p640x1536:94624

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p640x1536:94625

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p640x1536:94626

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p640x1536:94627

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p640x1536:94628

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p960x1024:216264

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p960x1024:216265

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p960x1024:216266

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p960x1024:216267

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p960x1024:216268

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p960x1024:216269

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p960x1024:216270

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p960x1024:216271

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p704x1472:94699

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p704x1472:94700

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p704x1472:94701

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p704x1472:94702

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p704x1472:94703

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p704x1472:94704

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p704x1472:94705

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

p704x1472:94706

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

r256_p1344x704:15577

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

r256_p1344x704:15578

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

r256_p1344x704:15579

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

r256_p1344x704:15580

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

r256_p1344x704:15581

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

r256_p1344x704:15582

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

r256_p1344x704:15583

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

r256_p1344x704:15584

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

r256_p896x1152:144131

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

r256_p896x1152:144132

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

r256_p896x1152:144133

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

r256_p896x1152:144134

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

r256_p896x1152:144135

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

r256_p896x1152:144136

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

r256_p896x1152:144137

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

r256_p896x1152:144138

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE

VAE_accuracy_test_image

Original
mdiffae_v1 (1 step)
Flux.1 VAE
Flux.2 VAE