LamTNguyen's picture
REPA representation probing: report, results, analysis, code
aaa6483 verified
|
Raw History Blame Contribute Delete
75.9 kB

Analysis tables

Teacher probe scores

teacher imagenet_cls/top1 latent_recon/r2 nyu_depth/delta1 nyu_depth/rmse rank/pr_pooled rank/pr_token
clip-vit-L 0.7365 0.1455 0.6040 0.7712 53.1939 4.2455
dinov2-vit-b 0.7278 0.3075 0.7480 0.5909 120.9584 225.5819
dinov3-vit-b16 0.7470 0.4302 0.8166 0.5082 145.2096 152.6689
dinov3-vit-h16plus 0.8276 0.3730 0.8040 0.5257 29.6966 67.2409
langpe-vit-g 0.7965 0.2765 0.7214 0.6254 74.7751 1.5402
mocov3-vit-b 0.5812 0.4245 0.6395 0.7113 92.8684 187.2330
pe-vit-g 0.8257 0.1881 0.5922 0.7946 38.6410 4.2698
webssl-vit-dino1b_full2b_224 0.7531 0.2992 0.8077 0.5300 56.4541 25.5436

Spearman: teacher score vs aligned-student score (8 teachers)

rho with bootstrap 95% CI (10k resamples of teachers) and exact two-sided permutation p (8! perms).

method task metric variant rho ci_lo ci_hi p_perm n
repa imagenet_cls top1 best(layer,s) -0.143 -0.973 0.747 0.752 8
repa imagenet_cls top1 best_layer@s=0.1 -0.143 -0.973 0.747 0.752 8
repa imagenet_cls top1 block8@s=0.1 -0.143 -0.973 0.747 0.752 8
repa imagenet_cls top1 best_layer@s=0.3 -0.143 -0.973 0.747 0.752 8
repa imagenet_cls top1 block8@s=0.3 -0.143 -0.973 0.747 0.752 8
repa imagenet_cls top1 best_layer@s=0.5 -0.156 -0.981 0.744 0.716 8
repa imagenet_cls top1 block8@s=0.5 -0.167 -1.000 0.737 0.703 8
repa imagenet_cls top1 best_layer@s=0.7 -0.167 -0.846 0.722 0.703 8
repa imagenet_cls top1 block8@s=0.7 0.048 -0.846 0.974 0.935 8
repa nyu_depth rmse best(layer,s) 0.595 -0.308 1.000 0.132 8
repa nyu_depth rmse best_layer@s=0.1 0.595 -0.308 1.000 0.132 8
repa nyu_depth rmse block8@s=0.1 0.524 -0.432 1.000 0.197 8
repa nyu_depth rmse best_layer@s=0.3 0.405 -0.729 1.000 0.327 8
repa nyu_depth rmse block8@s=0.3 0.452 -0.538 1.000 0.267 8
repa nyu_depth rmse best_layer@s=0.5 0.333 -0.737 0.926 0.428 8
repa nyu_depth rmse block8@s=0.5 0.381 -0.600 0.926 0.360 8
repa nyu_depth rmse best_layer@s=0.7 0.167 -0.775 0.974 0.703 8
repa nyu_depth rmse block8@s=0.7 0.071 -0.923 0.848 0.882 8
repa nyu_depth delta1 best(layer,s) 0.429 -0.615 0.973 0.299 8
repa nyu_depth delta1 best_layer@s=0.1 0.429 -0.615 0.973 0.299 8
repa nyu_depth delta1 block8@s=0.1 0.476 -0.468 1.000 0.243 8
repa nyu_depth delta1 best_layer@s=0.3 0.429 -0.615 0.973 0.299 8
repa nyu_depth delta1 block8@s=0.3 0.381 -0.600 0.946 0.360 8
repa nyu_depth delta1 best_layer@s=0.5 0.238 -1.000 0.899 0.582 8
repa nyu_depth delta1 block8@s=0.5 0.429 -0.615 0.973 0.299 8
repa nyu_depth delta1 best_layer@s=0.7 0.548 -0.333 1.000 0.171 8
repa nyu_depth delta1 block8@s=0.7 -0.643 -0.975 0.291 0.096 8
repa latent_recon r2 best(layer,s) -0.048 -0.842 0.923 0.935 8
repa latent_recon r2 best_layer@s=0.1 -0.048 -0.842 0.923 0.935 8
repa latent_recon r2 block8@s=0.1 -0.595 -0.975 0.233 0.132 8
repa latent_recon r2 best_layer@s=0.3 0.071 -0.696 0.846 0.882 8
repa latent_recon r2 block8@s=0.3 -0.357 -0.899 0.519 0.389 8
repa latent_recon r2 best_layer@s=0.5 0.405 -0.457 1.000 0.327 8
repa latent_recon r2 block8@s=0.5 -0.357 -0.899 0.519 0.389 8
repa latent_recon r2 best_layer@s=0.7 0.452 -0.333 1.000 0.267 8
repa latent_recon r2 block8@s=0.7 -0.310 -1.000 0.595 0.462 8
irepa imagenet_cls top1 best(layer,s) 0.190 -0.778 0.950 0.665 8
irepa imagenet_cls top1 best_layer@s=0.1 0.190 -0.778 0.950 0.665 8
irepa imagenet_cls top1 block8@s=0.1 0.190 -0.778 0.950 0.665 8
irepa imagenet_cls top1 best_layer@s=0.3 0.143 -0.842 0.905 0.752 8
irepa imagenet_cls top1 block8@s=0.3 0.143 -0.842 0.905 0.752 8
irepa imagenet_cls top1 best_layer@s=0.5 0.000 -0.926 0.795 1.000 8
irepa imagenet_cls top1 block8@s=0.5 0.000 -0.926 0.795 1.000 8
irepa imagenet_cls top1 best_layer@s=0.7 0.143 -0.737 1.000 0.752 8
irepa imagenet_cls top1 block8@s=0.7 0.143 -0.737 1.000 0.752 8
irepa nyu_depth rmse best(layer,s) 0.643 -0.139 1.000 0.096 8
irepa nyu_depth rmse best_layer@s=0.1 0.643 -0.139 1.000 0.096 8
irepa nyu_depth rmse block8@s=0.1 0.690 -0.114 1.000 0.069 8
irepa nyu_depth rmse best_layer@s=0.3 0.690 -0.114 1.000 0.069 8
irepa nyu_depth rmse block8@s=0.3 0.643 -0.215 1.000 0.096 8
irepa nyu_depth rmse best_layer@s=0.5 0.310 -0.973 0.926 0.462 8
irepa nyu_depth rmse block8@s=0.5 0.548 -0.407 1.000 0.171 8
irepa nyu_depth rmse best_layer@s=0.7 0.357 -0.615 1.000 0.389 8
irepa nyu_depth rmse block8@s=0.7 0.190 -1.000 0.899 0.665 8
irepa nyu_depth delta1 best(layer,s) 0.714 0.013 1.000 0.058 8
irepa nyu_depth delta1 best_layer@s=0.1 0.714 0.013 1.000 0.058 8
irepa nyu_depth delta1 block8@s=0.1 0.762 0.062 1.000 0.037 8
irepa nyu_depth delta1 best_layer@s=0.3 0.762 -0.062 1.000 0.037 8
irepa nyu_depth delta1 block8@s=0.3 0.643 -0.143 1.000 0.096 8
irepa nyu_depth delta1 best_layer@s=0.5 0.690 -0.139 1.000 0.069 8
irepa nyu_depth delta1 block8@s=0.5 0.619 -0.211 1.000 0.115 8
irepa nyu_depth delta1 best_layer@s=0.7 0.357 -0.781 0.899 0.389 8
irepa nyu_depth delta1 block8@s=0.7 0.405 -0.836 0.975 0.327 8
irepa latent_recon r2 best(layer,s) 0.667 -0.215 1.000 0.083 8
irepa latent_recon r2 best_layer@s=0.1 0.667 -0.215 1.000 0.083 8
irepa latent_recon r2 block8@s=0.1 0.405 -0.615 0.975 0.327 8
irepa latent_recon r2 best_layer@s=0.3 0.262 -0.595 0.926 0.536 8
irepa latent_recon r2 block8@s=0.3 -0.048 -0.836 0.753 0.935 8
irepa latent_recon r2 best_layer@s=0.5 0.190 -0.595 0.842 0.665 8
irepa latent_recon r2 block8@s=0.5 0.024 -1.000 0.622 0.977 8
irepa latent_recon r2 best_layer@s=0.7 0.286 -0.589 0.923 0.501 8
irepa latent_recon r2 block8@s=0.7 0.143 -1.000 0.641 0.752 8

Profile test: teacher contrast vs student contrast (controls for overall teacher strength)

method contrast variant rho ci_lo ci_hi p_perm n
repa nyu_depth/delta1 - imagenet_cls/top1 best(layer,s) 0.548 -0.333 1.000 0.171 8
repa nyu_depth/delta1 - imagenet_cls/top1 block8@s=0.1 0.405 -0.750 1.000 0.327 8
repa nyu_depth/delta1 - imagenet_cls/top1 block8@s=0.3 0.333 -0.842 1.000 0.428 8
repa nyu_depth/delta1 - imagenet_cls/top1 block8@s=0.5 0.429 -0.550 1.000 0.299 8
repa nyu_depth/delta1 - imagenet_cls/top1 block8@s=0.7 -0.238 -1.000 0.630 0.582 8
repa latent_recon/r2 - imagenet_cls/top1 best(layer,s) 0.143 -0.879 0.924 0.752 8
repa latent_recon/r2 - imagenet_cls/top1 block8@s=0.1 -0.143 -0.923 0.737 0.752 8
repa latent_recon/r2 - imagenet_cls/top1 block8@s=0.3 -0.143 -0.923 0.737 0.752 8
repa latent_recon/r2 - imagenet_cls/top1 block8@s=0.5 -0.143 -0.923 0.737 0.752 8
repa latent_recon/r2 - imagenet_cls/top1 block8@s=0.7 -0.048 -0.900 0.778 0.935 8
irepa nyu_depth/delta1 - imagenet_cls/top1 best(layer,s) 0.810 0.114 1.000 0.022 8
irepa nyu_depth/delta1 - imagenet_cls/top1 block8@s=0.1 0.762 0.089 1.000 0.037 8
irepa nyu_depth/delta1 - imagenet_cls/top1 block8@s=0.3 0.762 0.089 1.000 0.037 8
irepa nyu_depth/delta1 - imagenet_cls/top1 block8@s=0.5 0.738 -0.012 1.000 0.046 8
irepa nyu_depth/delta1 - imagenet_cls/top1 block8@s=0.7 0.095 -0.975 0.905 0.840 8
irepa latent_recon/r2 - imagenet_cls/top1 best(layer,s) 0.476 -0.400 0.975 0.243 8
irepa latent_recon/r2 - imagenet_cls/top1 block8@s=0.1 0.286 -0.722 0.973 0.501 8
irepa latent_recon/r2 - imagenet_cls/top1 block8@s=0.3 0.119 -0.842 1.000 0.793 8
irepa latent_recon/r2 - imagenet_cls/top1 block8@s=0.5 -0.071 -1.000 0.795 0.882 8
irepa latent_recon/r2 - imagenet_cls/top1 block8@s=0.7 -0.214 -1.000 0.671 0.619 8

Pairwise comparisons

comparison task metric cell a_mean a_where b_mean b_where diff sd_combined better verdict
REPA vs iREPA (dinov2-vit-b) imagenet_cls top1 best(layer,s) 0.5492 b8/s0.1 0.5827 b8/s0.1 -0.0335 0.0030 sit-xl-2-dinov2-vit-b-irepa conclusive
REPA vs iREPA (dinov2-vit-b) imagenet_cls top1 block8@s=0.1 0.5492 b8/s0.1 0.5827 b8/s0.1 -0.0335 0.0030 sit-xl-2-dinov2-vit-b-irepa conclusive
REPA vs iREPA (dinov2-vit-b) imagenet_cls top1 block8@s=0.3 0.5012 b8/s0.3 0.5377 b8/s0.3 -0.0365 0.0026 sit-xl-2-dinov2-vit-b-irepa conclusive
REPA vs iREPA (dinov2-vit-b) imagenet_cls top1 block8@s=0.5 0.3396 b8/s0.5 0.3641 b8/s0.5 -0.0245 0.0011 sit-xl-2-dinov2-vit-b-irepa conclusive
REPA vs iREPA (dinov2-vit-b) imagenet_cls top1 block8@s=0.7 0.1076 b8/s0.7 0.1042 b8/s0.7 0.0034 0.0020 sit-xl-2-dinov2-vit-b-repa inconclusive (<2 sd)
REPA vs iREPA (dinov2-vit-b) nyu_depth rmse best(layer,s) 0.6236 b12/s0.1 0.5993 b8/s0.1 0.0244 0.0076 sit-xl-2-dinov2-vit-b-irepa conclusive
REPA vs iREPA (dinov2-vit-b) nyu_depth rmse block8@s=0.1 0.6423 b8/s0.1 0.5993 b8/s0.1 0.0431 0.0019 sit-xl-2-dinov2-vit-b-irepa conclusive
REPA vs iREPA (dinov2-vit-b) nyu_depth rmse block8@s=0.3 0.6804 b8/s0.3 0.6378 b8/s0.3 0.0426 0.0011 sit-xl-2-dinov2-vit-b-irepa conclusive
REPA vs iREPA (dinov2-vit-b) nyu_depth rmse block8@s=0.5 0.7970 b8/s0.5 0.7863 b8/s0.5 0.0108 0.0078 sit-xl-2-dinov2-vit-b-irepa inconclusive (<2 sd)
REPA vs iREPA (dinov2-vit-b) nyu_depth rmse block8@s=0.7 0.9916 b8/s0.7 0.9946 b8/s0.7 -0.0030 0.0018 sit-xl-2-dinov2-vit-b-repa inconclusive (<2 sd)
REPA vs iREPA (dinov2-vit-b) nyu_depth delta1 best(layer,s) 0.7130 b12/s0.1 0.7342 b8/s0.1 -0.0212 0.0051 sit-xl-2-dinov2-vit-b-irepa conclusive
REPA vs iREPA (dinov2-vit-b) nyu_depth delta1 block8@s=0.1 0.6967 b8/s0.1 0.7342 b8/s0.1 -0.0376 0.0013 sit-xl-2-dinov2-vit-b-irepa conclusive
REPA vs iREPA (dinov2-vit-b) nyu_depth delta1 block8@s=0.3 0.6666 b8/s0.3 0.7017 b8/s0.3 -0.0351 0.0006 sit-xl-2-dinov2-vit-b-irepa conclusive
REPA vs iREPA (dinov2-vit-b) nyu_depth delta1 block8@s=0.5 0.5854 b8/s0.5 0.5945 b8/s0.5 -0.0090 0.0048 sit-xl-2-dinov2-vit-b-irepa inconclusive (<2 sd)
REPA vs iREPA (dinov2-vit-b) nyu_depth delta1 block8@s=0.7 0.4742 b8/s0.7 0.4738 b8/s0.7 0.0003 0.0008 sit-xl-2-dinov2-vit-b-repa inconclusive (<2 sd)
REPA vs iREPA (dinov2-vit-b) latent_recon r2 best(layer,s) 0.9842 b24/s0.1 0.9843 b28/s0.1 -0.0001 0.0000 sit-xl-2-dinov2-vit-b-irepa conclusive
REPA vs iREPA (dinov2-vit-b) latent_recon r2 block8@s=0.1 0.9812 b8/s0.1 0.9799 b8/s0.1 0.0013 0.0000 sit-xl-2-dinov2-vit-b-repa conclusive
REPA vs iREPA (dinov2-vit-b) latent_recon r2 block8@s=0.3 0.8453 b8/s0.3 0.8364 b8/s0.3 0.0089 0.0000 sit-xl-2-dinov2-vit-b-repa conclusive
REPA vs iREPA (dinov2-vit-b) latent_recon r2 block8@s=0.5 0.6452 b8/s0.5 0.6348 b8/s0.5 0.0104 0.0000 sit-xl-2-dinov2-vit-b-repa conclusive
REPA vs iREPA (dinov2-vit-b) latent_recon r2 block8@s=0.7 0.4565 b8/s0.7 0.4520 b8/s0.7 0.0046 0.0000 sit-xl-2-dinov2-vit-b-repa conclusive
REPA vs iREPA (dinov2-vit-b) cknna cknna_teacher best(layer,s) 0.4955 b8/s0.1 0.6284 b8/s0.1 -0.1329 0.0000 sit-xl-2-dinov2-vit-b-irepa n/a (no seed std)
REPA vs iREPA (dinov2-vit-b) cknna cknna_teacher block8@s=0.1 0.4955 b8/s0.1 0.6284 b8/s0.1 -0.1329 0.0000 sit-xl-2-dinov2-vit-b-irepa n/a (no seed std)
REPA vs iREPA (dinov2-vit-b) cknna cknna_teacher block8@s=0.3 0.4522 b8/s0.3 0.5602 b8/s0.3 -0.1080 0.0000 sit-xl-2-dinov2-vit-b-irepa n/a (no seed std)
REPA vs iREPA (dinov2-vit-b) cknna cknna_teacher block8@s=0.5 0.2882 b8/s0.5 0.3408 b8/s0.5 -0.0526 0.0000 sit-xl-2-dinov2-vit-b-irepa n/a (no seed std)
REPA vs iREPA (dinov2-vit-b) cknna cknna_teacher block8@s=0.7 0.0697 b8/s0.7 0.0745 b8/s0.7 -0.0048 0.0000 sit-xl-2-dinov2-vit-b-irepa n/a (no seed std)
REPA vs iREPA (dinov2-vit-b) cknna cknna_dinov2 best(layer,s) 0.4955 b8/s0.1 0.6284 b8/s0.1 -0.1329 0.0000 sit-xl-2-dinov2-vit-b-irepa n/a (no seed std)
REPA vs iREPA (dinov2-vit-b) cknna cknna_dinov2 block8@s=0.1 0.4955 b8/s0.1 0.6284 b8/s0.1 -0.1329 0.0000 sit-xl-2-dinov2-vit-b-irepa n/a (no seed std)
REPA vs iREPA (dinov2-vit-b) cknna cknna_dinov2 block8@s=0.3 0.4522 b8/s0.3 0.5602 b8/s0.3 -0.1080 0.0000 sit-xl-2-dinov2-vit-b-irepa n/a (no seed std)
REPA vs iREPA (dinov2-vit-b) cknna cknna_dinov2 block8@s=0.5 0.2882 b8/s0.5 0.3408 b8/s0.5 -0.0526 0.0000 sit-xl-2-dinov2-vit-b-irepa n/a (no seed std)
REPA vs iREPA (dinov2-vit-b) cknna cknna_dinov2 block8@s=0.7 0.0697 b8/s0.7 0.0745 b8/s0.7 -0.0048 0.0000 sit-xl-2-dinov2-vit-b-irepa n/a (no seed std)
REPA vs iREPA (dinov2-vit-b) rank pr_token best(layer,s) 208.8756 b8/s0.3 373.5329 b8/s0.1 -164.6573 0.0000 n/a (no seed std)
REPA vs iREPA (dinov2-vit-b) rank pr_token block8@s=0.1 190.6034 b8/s0.1 373.5329 b8/s0.1 -182.9294 0.0000 n/a (no seed std)
REPA vs iREPA (dinov2-vit-b) rank pr_token block8@s=0.3 208.8756 b8/s0.3 369.6074 b8/s0.3 -160.7318 0.0000 n/a (no seed std)
REPA vs iREPA (dinov2-vit-b) rank pr_token block8@s=0.5 184.5681 b8/s0.5 308.7439 b8/s0.5 -124.1758 0.0000 n/a (no seed std)
REPA vs iREPA (dinov2-vit-b) rank pr_token block8@s=0.7 113.6005 b8/s0.7 168.5247 b8/s0.7 -54.9242 0.0000 n/a (no seed std)
REPA vs iREPA (dinov3-vit-b16) imagenet_cls top1 best(layer,s) 0.5426 b8/s0.1 0.5944 b8/s0.1 -0.0518 0.0035 sit-xl-2-dinov3-vit-b16-irepa conclusive
REPA vs iREPA (dinov3-vit-b16) imagenet_cls top1 block8@s=0.1 0.5426 b8/s0.1 0.5944 b8/s0.1 -0.0518 0.0035 sit-xl-2-dinov3-vit-b16-irepa conclusive
REPA vs iREPA (dinov3-vit-b16) imagenet_cls top1 block8@s=0.3 0.4948 b8/s0.3 0.5465 b8/s0.3 -0.0517 0.0026 sit-xl-2-dinov3-vit-b16-irepa conclusive
REPA vs iREPA (dinov3-vit-b16) imagenet_cls top1 block8@s=0.5 0.3291 b8/s0.5 0.3656 b8/s0.5 -0.0365 0.0025 sit-xl-2-dinov3-vit-b16-irepa conclusive
REPA vs iREPA (dinov3-vit-b16) imagenet_cls top1 block8@s=0.7 0.1027 b8/s0.7 0.1068 b8/s0.7 -0.0041 0.0021 sit-xl-2-dinov3-vit-b16-irepa conclusive
REPA vs iREPA (dinov3-vit-b16) nyu_depth rmse best(layer,s) 0.6200 b12/s0.1 0.5809 b12/s0.1 0.0391 0.0037 sit-xl-2-dinov3-vit-b16-irepa conclusive
REPA vs iREPA (dinov3-vit-b16) nyu_depth rmse block8@s=0.1 0.6358 b8/s0.1 0.5913 b8/s0.1 0.0444 0.0010 sit-xl-2-dinov3-vit-b16-irepa conclusive
REPA vs iREPA (dinov3-vit-b16) nyu_depth rmse block8@s=0.3 0.6868 b8/s0.3 0.6323 b8/s0.3 0.0545 0.0012 sit-xl-2-dinov3-vit-b16-irepa conclusive
REPA vs iREPA (dinov3-vit-b16) nyu_depth rmse block8@s=0.5 0.8027 b8/s0.5 0.7817 b8/s0.5 0.0210 0.0086 sit-xl-2-dinov3-vit-b16-irepa conclusive
REPA vs iREPA (dinov3-vit-b16) nyu_depth rmse block8@s=0.7 0.9918 b8/s0.7 0.9908 b8/s0.7 0.0011 0.0013 sit-xl-2-dinov3-vit-b16-irepa inconclusive (<2 sd)
REPA vs iREPA (dinov3-vit-b16) nyu_depth delta1 best(layer,s) 0.7093 b12/s0.1 0.7417 b12/s0.1 -0.0324 0.0020 sit-xl-2-dinov3-vit-b16-irepa conclusive
REPA vs iREPA (dinov3-vit-b16) nyu_depth delta1 block8@s=0.1 0.6948 b8/s0.1 0.7345 b8/s0.1 -0.0397 0.0009 sit-xl-2-dinov3-vit-b16-irepa conclusive
REPA vs iREPA (dinov3-vit-b16) nyu_depth delta1 block8@s=0.3 0.6538 b8/s0.3 0.6980 b8/s0.3 -0.0441 0.0010 sit-xl-2-dinov3-vit-b16-irepa conclusive
REPA vs iREPA (dinov3-vit-b16) nyu_depth delta1 block8@s=0.5 0.5760 b8/s0.5 0.5942 b8/s0.5 -0.0182 0.0088 sit-xl-2-dinov3-vit-b16-irepa conclusive
REPA vs iREPA (dinov3-vit-b16) nyu_depth delta1 block8@s=0.7 0.4686 b8/s0.7 0.4752 b8/s0.7 -0.0066 0.0012 sit-xl-2-dinov3-vit-b16-irepa conclusive
REPA vs iREPA (dinov3-vit-b16) latent_recon r2 best(layer,s) 0.9838 b24/s0.1 0.9835 b24/s0.1 0.0003 0.0000 sit-xl-2-dinov3-vit-b16-repa conclusive
REPA vs iREPA (dinov3-vit-b16) latent_recon r2 block8@s=0.1 0.9812 b8/s0.1 0.9792 b8/s0.1 0.0019 0.0000 sit-xl-2-dinov3-vit-b16-repa conclusive
REPA vs iREPA (dinov3-vit-b16) latent_recon r2 block8@s=0.3 0.8458 b8/s0.3 0.8379 b8/s0.3 0.0079 0.0000 sit-xl-2-dinov3-vit-b16-repa conclusive
REPA vs iREPA (dinov3-vit-b16) latent_recon r2 block8@s=0.5 0.6476 b8/s0.5 0.6412 b8/s0.5 0.0064 0.0000 sit-xl-2-dinov3-vit-b16-repa conclusive
REPA vs iREPA (dinov3-vit-b16) latent_recon r2 block8@s=0.7 0.4579 b8/s0.7 0.4569 b8/s0.7 0.0010 0.0000 sit-xl-2-dinov3-vit-b16-repa conclusive
REPA vs iREPA (dinov3-vit-b16) cknna cknna_teacher best(layer,s) 0.4306 b8/s0.1 0.5611 b8/s0.1 -0.1305 0.0000 sit-xl-2-dinov3-vit-b16-irepa n/a (no seed std)
REPA vs iREPA (dinov3-vit-b16) cknna cknna_teacher block8@s=0.1 0.4306 b8/s0.1 0.5611 b8/s0.1 -0.1305 0.0000 sit-xl-2-dinov3-vit-b16-irepa n/a (no seed std)
REPA vs iREPA (dinov3-vit-b16) cknna cknna_teacher block8@s=0.3 0.3878 b8/s0.3 0.4903 b8/s0.3 -0.1025 0.0000 sit-xl-2-dinov3-vit-b16-irepa n/a (no seed std)
REPA vs iREPA (dinov3-vit-b16) cknna cknna_teacher block8@s=0.5 0.2353 b8/s0.5 0.2822 b8/s0.5 -0.0470 0.0000 sit-xl-2-dinov3-vit-b16-irepa n/a (no seed std)
REPA vs iREPA (dinov3-vit-b16) cknna cknna_teacher block8@s=0.7 0.0509 b8/s0.7 0.0564 b8/s0.7 -0.0055 0.0000 sit-xl-2-dinov3-vit-b16-irepa n/a (no seed std)
REPA vs iREPA (dinov3-vit-b16) rank pr_token best(layer,s) 156.2727 b8/s0.3 298.5637 b8/s0.1 -142.2909 0.0000 n/a (no seed std)
REPA vs iREPA (dinov3-vit-b16) rank pr_token block8@s=0.1 132.4590 b8/s0.1 298.5637 b8/s0.1 -166.1046 0.0000 n/a (no seed std)
REPA vs iREPA (dinov3-vit-b16) rank pr_token block8@s=0.3 156.2727 b8/s0.3 275.9211 b8/s0.3 -119.6484 0.0000 n/a (no seed std)
REPA vs iREPA (dinov3-vit-b16) rank pr_token block8@s=0.5 143.6176 b8/s0.5 214.0842 b8/s0.5 -70.4665 0.0000 n/a (no seed std)
REPA vs iREPA (dinov3-vit-b16) rank pr_token block8@s=0.7 89.3007 b8/s0.7 109.6463 b8/s0.7 -20.3455 0.0000 n/a (no seed std)
REPA vs iREPA (dinov3-vit-h16plus) imagenet_cls top1 best(layer,s) 0.4863 b8/s0.1 0.5610 b8/s0.1 -0.0747 0.0029 sit-xl-2-dinov3-vit-h16plus-irepa conclusive
REPA vs iREPA (dinov3-vit-h16plus) imagenet_cls top1 block8@s=0.1 0.4863 b8/s0.1 0.5610 b8/s0.1 -0.0747 0.0029 sit-xl-2-dinov3-vit-h16plus-irepa conclusive
REPA vs iREPA (dinov3-vit-h16plus) imagenet_cls top1 block8@s=0.3 0.4404 b8/s0.3 0.5110 b8/s0.3 -0.0706 0.0011 sit-xl-2-dinov3-vit-h16plus-irepa conclusive
REPA vs iREPA (dinov3-vit-h16plus) imagenet_cls top1 block8@s=0.5 0.2918 b8/s0.5 0.3439 b8/s0.5 -0.0521 0.0025 sit-xl-2-dinov3-vit-h16plus-irepa conclusive
REPA vs iREPA (dinov3-vit-h16plus) imagenet_cls top1 block8@s=0.7 0.0933 b8/s0.7 0.1016 b8/s0.7 -0.0083 0.0030 sit-xl-2-dinov3-vit-h16plus-irepa conclusive
REPA vs iREPA (dinov3-vit-h16plus) nyu_depth rmse best(layer,s) 0.6661 b12/s0.1 0.6239 b12/s0.1 0.0422 0.0057 sit-xl-2-dinov3-vit-h16plus-irepa conclusive
REPA vs iREPA (dinov3-vit-h16plus) nyu_depth rmse block8@s=0.1 0.6931 b8/s0.1 0.6528 b8/s0.1 0.0404 0.0028 sit-xl-2-dinov3-vit-h16plus-irepa conclusive
REPA vs iREPA (dinov3-vit-h16plus) nyu_depth rmse block8@s=0.3 0.7320 b8/s0.3 0.6904 b8/s0.3 0.0416 0.0046 sit-xl-2-dinov3-vit-h16plus-irepa conclusive
REPA vs iREPA (dinov3-vit-h16plus) nyu_depth rmse block8@s=0.5 0.8366 b8/s0.5 0.8273 b8/s0.5 0.0093 0.0027 sit-xl-2-dinov3-vit-h16plus-irepa conclusive
REPA vs iREPA (dinov3-vit-h16plus) nyu_depth rmse block8@s=0.7 0.9912 b8/s0.7 0.9988 b8/s0.7 -0.0076 0.0122 sit-xl-2-dinov3-vit-h16plus-repa inconclusive (<2 sd)
REPA vs iREPA (dinov3-vit-h16plus) nyu_depth delta1 best(layer,s) 0.6693 b12/s0.1 0.7093 b12/s0.1 -0.0400 0.0040 sit-xl-2-dinov3-vit-h16plus-irepa conclusive
REPA vs iREPA (dinov3-vit-h16plus) nyu_depth delta1 block8@s=0.1 0.6489 b8/s0.1 0.6856 b8/s0.1 -0.0368 0.0031 sit-xl-2-dinov3-vit-h16plus-irepa conclusive
REPA vs iREPA (dinov3-vit-h16plus) nyu_depth delta1 block8@s=0.3 0.6180 b8/s0.3 0.6546 b8/s0.3 -0.0366 0.0036 sit-xl-2-dinov3-vit-h16plus-irepa conclusive
REPA vs iREPA (dinov3-vit-h16plus) nyu_depth delta1 block8@s=0.5 0.5544 b8/s0.5 0.5656 b8/s0.5 -0.0112 0.0021 sit-xl-2-dinov3-vit-h16plus-irepa conclusive
REPA vs iREPA (dinov3-vit-h16plus) nyu_depth delta1 block8@s=0.7 0.4709 b8/s0.7 0.4733 b8/s0.7 -0.0023 0.0068 sit-xl-2-dinov3-vit-h16plus-irepa inconclusive (<2 sd)
REPA vs iREPA (dinov3-vit-h16plus) latent_recon r2 best(layer,s) 0.9839 b24/s0.1 0.9840 b28/s0.1 -0.0000 0.0000 sit-xl-2-dinov3-vit-h16plus-irepa conclusive
REPA vs iREPA (dinov3-vit-h16plus) latent_recon r2 block8@s=0.1 0.9818 b8/s0.1 0.9794 b8/s0.1 0.0024 0.0000 sit-xl-2-dinov3-vit-h16plus-repa conclusive
REPA vs iREPA (dinov3-vit-h16plus) latent_recon r2 block8@s=0.3 0.8456 b8/s0.3 0.8300 b8/s0.3 0.0156 0.0000 sit-xl-2-dinov3-vit-h16plus-repa conclusive
REPA vs iREPA (dinov3-vit-h16plus) latent_recon r2 block8@s=0.5 0.6458 b8/s0.5 0.6279 b8/s0.5 0.0179 0.0000 sit-xl-2-dinov3-vit-h16plus-repa conclusive
REPA vs iREPA (dinov3-vit-h16plus) latent_recon r2 block8@s=0.7 0.4562 b8/s0.7 0.4499 b8/s0.7 0.0063 0.0000 sit-xl-2-dinov3-vit-h16plus-repa conclusive
REPA vs iREPA (dinov3-vit-h16plus) cknna cknna_teacher best(layer,s) 0.2522 b8/s0.1 0.3643 b8/s0.1 -0.1121 0.0000 sit-xl-2-dinov3-vit-h16plus-irepa n/a (no seed std)
REPA vs iREPA (dinov3-vit-h16plus) cknna cknna_teacher block8@s=0.1 0.2522 b8/s0.1 0.3643 b8/s0.1 -0.1121 0.0000 sit-xl-2-dinov3-vit-h16plus-irepa n/a (no seed std)
REPA vs iREPA (dinov3-vit-h16plus) cknna cknna_teacher block8@s=0.3 0.2234 b8/s0.3 0.3144 b8/s0.3 -0.0910 0.0000 sit-xl-2-dinov3-vit-h16plus-irepa n/a (no seed std)
REPA vs iREPA (dinov3-vit-h16plus) cknna cknna_teacher block8@s=0.5 0.1298 b8/s0.5 0.1842 b8/s0.5 -0.0543 0.0000 sit-xl-2-dinov3-vit-h16plus-irepa n/a (no seed std)
REPA vs iREPA (dinov3-vit-h16plus) cknna cknna_teacher block8@s=0.7 0.0313 b8/s0.7 0.0385 b8/s0.7 -0.0072 0.0000 sit-xl-2-dinov3-vit-h16plus-irepa n/a (no seed std)
REPA vs iREPA (dinov3-vit-h16plus) rank pr_token best(layer,s) 93.7348 b28/s0.1 239.4351 b8/s0.1 -145.7002 0.0000 n/a (no seed std)
REPA vs iREPA (dinov3-vit-h16plus) rank pr_token block8@s=0.1 74.7451 b8/s0.1 239.4351 b8/s0.1 -164.6900 0.0000 n/a (no seed std)
REPA vs iREPA (dinov3-vit-h16plus) rank pr_token block8@s=0.3 93.2526 b8/s0.3 221.7327 b8/s0.3 -128.4802 0.0000 n/a (no seed std)
REPA vs iREPA (dinov3-vit-h16plus) rank pr_token block8@s=0.5 91.2506 b8/s0.5 174.6181 b8/s0.5 -83.3676 0.0000 n/a (no seed std)
REPA vs iREPA (dinov3-vit-h16plus) rank pr_token block8@s=0.7 64.9301 b8/s0.7 93.5096 b8/s0.7 -28.5795 0.0000 n/a (no seed std)
REPA vs iREPA (clip-vit-L) imagenet_cls top1 best(layer,s) 0.4945 b8/s0.1 0.5647 b8/s0.1 -0.0701 0.0037 sit-xl-2-clip-vit-L-irepa conclusive
REPA vs iREPA (clip-vit-L) imagenet_cls top1 block8@s=0.1 0.4945 b8/s0.1 0.5647 b8/s0.1 -0.0701 0.0037 sit-xl-2-clip-vit-L-irepa conclusive
REPA vs iREPA (clip-vit-L) imagenet_cls top1 block8@s=0.3 0.4462 b8/s0.3 0.5183 b8/s0.3 -0.0720 0.0039 sit-xl-2-clip-vit-L-irepa conclusive
REPA vs iREPA (clip-vit-L) imagenet_cls top1 block8@s=0.5 0.2942 b8/s0.5 0.3425 b8/s0.5 -0.0483 0.0023 sit-xl-2-clip-vit-L-irepa conclusive
REPA vs iREPA (clip-vit-L) imagenet_cls top1 block8@s=0.7 0.0918 b8/s0.7 0.0969 b8/s0.7 -0.0051 0.0020 sit-xl-2-clip-vit-L-irepa conclusive
REPA vs iREPA (clip-vit-L) nyu_depth rmse best(layer,s) 0.6601 b12/s0.1 0.6115 b12/s0.1 0.0486 0.0035 sit-xl-2-clip-vit-L-irepa conclusive
REPA vs iREPA (clip-vit-L) nyu_depth rmse block8@s=0.1 0.6653 b8/s0.1 0.6426 b8/s0.1 0.0226 0.0027 sit-xl-2-clip-vit-L-irepa conclusive
REPA vs iREPA (clip-vit-L) nyu_depth rmse block8@s=0.3 0.6989 b8/s0.3 0.6705 b8/s0.3 0.0285 0.0025 sit-xl-2-clip-vit-L-irepa conclusive
REPA vs iREPA (clip-vit-L) nyu_depth rmse block8@s=0.5 0.8271 b8/s0.5 0.8023 b8/s0.5 0.0247 0.0019 sit-xl-2-clip-vit-L-irepa conclusive
REPA vs iREPA (clip-vit-L) nyu_depth rmse block8@s=0.7 0.9945 b8/s0.7 0.9930 b8/s0.7 0.0015 0.0015 sit-xl-2-clip-vit-L-irepa inconclusive (<2 sd)
REPA vs iREPA (clip-vit-L) nyu_depth delta1 best(layer,s) 0.6846 b12/s0.1 0.7200 b12/s0.1 -0.0354 0.0015 sit-xl-2-clip-vit-L-irepa conclusive
REPA vs iREPA (clip-vit-L) nyu_depth delta1 block8@s=0.1 0.6760 b8/s0.1 0.6916 b8/s0.1 -0.0156 0.0012 sit-xl-2-clip-vit-L-irepa conclusive
REPA vs iREPA (clip-vit-L) nyu_depth delta1 block8@s=0.3 0.6502 b8/s0.3 0.6698 b8/s0.3 -0.0196 0.0017 sit-xl-2-clip-vit-L-irepa conclusive
REPA vs iREPA (clip-vit-L) nyu_depth delta1 block8@s=0.5 0.5605 b8/s0.5 0.5762 b8/s0.5 -0.0157 0.0008 sit-xl-2-clip-vit-L-irepa conclusive
REPA vs iREPA (clip-vit-L) nyu_depth delta1 block8@s=0.7 0.4727 b8/s0.7 0.4744 b8/s0.7 -0.0017 0.0007 sit-xl-2-clip-vit-L-irepa conclusive
REPA vs iREPA (clip-vit-L) latent_recon r2 best(layer,s) 0.9841 b24/s0.1 0.9833 b24/s0.1 0.0008 0.0000 sit-xl-2-clip-vit-L-repa conclusive
REPA vs iREPA (clip-vit-L) latent_recon r2 block8@s=0.1 0.9822 b8/s0.1 0.9788 b8/s0.1 0.0035 0.0000 sit-xl-2-clip-vit-L-repa conclusive
REPA vs iREPA (clip-vit-L) latent_recon r2 block8@s=0.3 0.8506 b8/s0.3 0.8379 b8/s0.3 0.0127 0.0000 sit-xl-2-clip-vit-L-repa conclusive
REPA vs iREPA (clip-vit-L) latent_recon r2 block8@s=0.5 0.6523 b8/s0.5 0.6400 b8/s0.5 0.0124 0.0000 sit-xl-2-clip-vit-L-repa conclusive
REPA vs iREPA (clip-vit-L) latent_recon r2 block8@s=0.7 0.4593 b8/s0.7 0.4557 b8/s0.7 0.0036 0.0000 sit-xl-2-clip-vit-L-repa conclusive
REPA vs iREPA (clip-vit-L) cknna cknna_teacher best(layer,s) 0.2941 b8/s0.1 0.4148 b8/s0.1 -0.1206 0.0000 sit-xl-2-clip-vit-L-irepa n/a (no seed std)
REPA vs iREPA (clip-vit-L) cknna cknna_teacher block8@s=0.1 0.2941 b8/s0.1 0.4148 b8/s0.1 -0.1206 0.0000 sit-xl-2-clip-vit-L-irepa n/a (no seed std)
REPA vs iREPA (clip-vit-L) cknna cknna_teacher block8@s=0.3 0.2693 b8/s0.3 0.3632 b8/s0.3 -0.0939 0.0000 sit-xl-2-clip-vit-L-irepa n/a (no seed std)
REPA vs iREPA (clip-vit-L) cknna cknna_teacher block8@s=0.5 0.1718 b8/s0.5 0.2132 b8/s0.5 -0.0414 0.0000 sit-xl-2-clip-vit-L-irepa n/a (no seed std)
REPA vs iREPA (clip-vit-L) cknna cknna_teacher block8@s=0.7 0.0468 b8/s0.7 0.0504 b8/s0.7 -0.0036 0.0000 sit-xl-2-clip-vit-L-irepa n/a (no seed std)
REPA vs iREPA (clip-vit-L) rank pr_token best(layer,s) 108.4278 b8/s0.3 171.1534 b8/s0.1 -62.7256 0.0000 n/a (no seed std)
REPA vs iREPA (clip-vit-L) rank pr_token block8@s=0.1 89.5579 b8/s0.1 171.1534 b8/s0.1 -81.5955 0.0000 n/a (no seed std)
REPA vs iREPA (clip-vit-L) rank pr_token block8@s=0.3 108.4278 b8/s0.3 164.2708 b8/s0.3 -55.8430 0.0000 n/a (no seed std)
REPA vs iREPA (clip-vit-L) rank pr_token block8@s=0.5 99.6180 b8/s0.5 122.1815 b8/s0.5 -22.5635 0.0000 n/a (no seed std)
REPA vs iREPA (clip-vit-L) rank pr_token block8@s=0.7 67.1628 b8/s0.7 56.7176 b8/s0.7 10.4451 0.0000 n/a (no seed std)
REPA vs iREPA (mocov3-vit-b) imagenet_cls top1 best(layer,s) 0.4045 b12/s0.1 0.4761 b8/s0.1 -0.0717 0.0034 sit-xl-2-mocov3-vit-b-irepa conclusive
REPA vs iREPA (mocov3-vit-b) imagenet_cls top1 block8@s=0.1 0.4037 b8/s0.1 0.4761 b8/s0.1 -0.0724 0.0042 sit-xl-2-mocov3-vit-b-irepa conclusive
REPA vs iREPA (mocov3-vit-b) imagenet_cls top1 block8@s=0.3 0.3666 b8/s0.3 0.4395 b8/s0.3 -0.0729 0.0018 sit-xl-2-mocov3-vit-b-irepa conclusive
REPA vs iREPA (mocov3-vit-b) imagenet_cls top1 block8@s=0.5 0.2484 b8/s0.5 0.3025 b8/s0.5 -0.0541 0.0012 sit-xl-2-mocov3-vit-b-irepa conclusive
REPA vs iREPA (mocov3-vit-b) imagenet_cls top1 block8@s=0.7 0.0800 b8/s0.7 0.0912 b8/s0.7 -0.0112 0.0035 sit-xl-2-mocov3-vit-b-irepa conclusive
REPA vs iREPA (mocov3-vit-b) nyu_depth rmse best(layer,s) 0.6863 b16/s0.1 0.6438 b12/s0.1 0.0425 0.0052 sit-xl-2-mocov3-vit-b-irepa conclusive
REPA vs iREPA (mocov3-vit-b) nyu_depth rmse block8@s=0.1 0.7009 b8/s0.1 0.6681 b8/s0.1 0.0328 0.0073 sit-xl-2-mocov3-vit-b-irepa conclusive
REPA vs iREPA (mocov3-vit-b) nyu_depth rmse block8@s=0.3 0.7389 b8/s0.3 0.6939 b8/s0.3 0.0450 0.0071 sit-xl-2-mocov3-vit-b-irepa conclusive
REPA vs iREPA (mocov3-vit-b) nyu_depth rmse block8@s=0.5 0.8444 b8/s0.5 0.8092 b8/s0.5 0.0353 0.0044 sit-xl-2-mocov3-vit-b-irepa conclusive
REPA vs iREPA (mocov3-vit-b) nyu_depth rmse block8@s=0.7 0.9892 b8/s0.7 0.9937 b8/s0.7 -0.0045 0.0150 sit-xl-2-mocov3-vit-b-repa inconclusive (<2 sd)
REPA vs iREPA (mocov3-vit-b) nyu_depth delta1 best(layer,s) 0.6542 b16/s0.1 0.6913 b12/s0.1 -0.0371 0.0033 sit-xl-2-mocov3-vit-b-irepa conclusive
REPA vs iREPA (mocov3-vit-b) nyu_depth delta1 block8@s=0.1 0.6412 b8/s0.1 0.6712 b8/s0.1 -0.0300 0.0057 sit-xl-2-mocov3-vit-b-irepa conclusive
REPA vs iREPA (mocov3-vit-b) nyu_depth delta1 block8@s=0.3 0.6180 b8/s0.3 0.6533 b8/s0.3 -0.0352 0.0056 sit-xl-2-mocov3-vit-b-irepa conclusive
REPA vs iREPA (mocov3-vit-b) nyu_depth delta1 block8@s=0.5 0.5475 b8/s0.5 0.5693 b8/s0.5 -0.0218 0.0061 sit-xl-2-mocov3-vit-b-irepa conclusive
REPA vs iREPA (mocov3-vit-b) nyu_depth delta1 block8@s=0.7 0.4738 b8/s0.7 0.4695 b8/s0.7 0.0044 0.0114 sit-xl-2-mocov3-vit-b-repa inconclusive (<2 sd)
REPA vs iREPA (mocov3-vit-b) latent_recon r2 best(layer,s) 0.9841 b28/s0.1 0.9842 b24/s0.1 -0.0001 0.0000 sit-xl-2-mocov3-vit-b-irepa conclusive
REPA vs iREPA (mocov3-vit-b) latent_recon r2 block8@s=0.1 0.9823 b8/s0.1 0.9809 b8/s0.1 0.0014 0.0000 sit-xl-2-mocov3-vit-b-repa conclusive
REPA vs iREPA (mocov3-vit-b) latent_recon r2 block8@s=0.3 0.8529 b8/s0.3 0.8479 b8/s0.3 0.0050 0.0000 sit-xl-2-mocov3-vit-b-repa conclusive
REPA vs iREPA (mocov3-vit-b) latent_recon r2 block8@s=0.5 0.6565 b8/s0.5 0.6538 b8/s0.5 0.0027 0.0000 sit-xl-2-mocov3-vit-b-repa conclusive
REPA vs iREPA (mocov3-vit-b) latent_recon r2 block8@s=0.7 0.4621 b8/s0.7 0.4635 b8/s0.7 -0.0014 0.0000 sit-xl-2-mocov3-vit-b-irepa conclusive
REPA vs iREPA (mocov3-vit-b) cknna cknna_teacher best(layer,s) 0.2975 b8/s0.1 0.4860 b8/s0.1 -0.1885 0.0000 sit-xl-2-mocov3-vit-b-irepa n/a (no seed std)
REPA vs iREPA (mocov3-vit-b) cknna cknna_teacher block8@s=0.1 0.2975 b8/s0.1 0.4860 b8/s0.1 -0.1885 0.0000 sit-xl-2-mocov3-vit-b-irepa n/a (no seed std)
REPA vs iREPA (mocov3-vit-b) cknna cknna_teacher block8@s=0.3 0.2813 b8/s0.3 0.4430 b8/s0.3 -0.1616 0.0000 sit-xl-2-mocov3-vit-b-irepa n/a (no seed std)
REPA vs iREPA (mocov3-vit-b) cknna cknna_teacher block8@s=0.5 0.1809 b8/s0.5 0.2897 b8/s0.5 -0.1088 0.0000 sit-xl-2-mocov3-vit-b-irepa n/a (no seed std)
REPA vs iREPA (mocov3-vit-b) cknna cknna_teacher block8@s=0.7 0.0596 b8/s0.7 0.0779 b8/s0.7 -0.0183 0.0000 sit-xl-2-mocov3-vit-b-irepa n/a (no seed std)
REPA vs iREPA (mocov3-vit-b) rank pr_token best(layer,s) 103.1277 b28/s0.1 324.2182 b8/s0.1 -221.0904 0.0000 n/a (no seed std)
REPA vs iREPA (mocov3-vit-b) rank pr_token block8@s=0.1 61.7894 b8/s0.1 324.2182 b8/s0.1 -262.4287 0.0000 n/a (no seed std)
REPA vs iREPA (mocov3-vit-b) rank pr_token block8@s=0.3 75.2855 b8/s0.3 305.2310 b8/s0.3 -229.9455 0.0000 n/a (no seed std)
REPA vs iREPA (mocov3-vit-b) rank pr_token block8@s=0.5 71.3629 b8/s0.5 240.1996 b8/s0.5 -168.8366 0.0000 n/a (no seed std)
REPA vs iREPA (mocov3-vit-b) rank pr_token block8@s=0.7 53.4480 b8/s0.7 130.8607 b8/s0.7 -77.4127 0.0000 n/a (no seed std)
REPA vs iREPA (webssl-vit-dino1b_full2b_224) imagenet_cls top1 best(layer,s) 0.4582 b8/s0.1 0.5889 b8/s0.1 -0.1308 0.0035 sit-xl-2-webssl-vit-dino1b_full2b_224-irepa conclusive
REPA vs iREPA (webssl-vit-dino1b_full2b_224) imagenet_cls top1 block8@s=0.1 0.4582 b8/s0.1 0.5889 b8/s0.1 -0.1308 0.0035 sit-xl-2-webssl-vit-dino1b_full2b_224-irepa conclusive
REPA vs iREPA (webssl-vit-dino1b_full2b_224) imagenet_cls top1 block8@s=0.3 0.4121 b8/s0.3 0.5383 b8/s0.3 -0.1262 0.0047 sit-xl-2-webssl-vit-dino1b_full2b_224-irepa conclusive
REPA vs iREPA (webssl-vit-dino1b_full2b_224) imagenet_cls top1 block8@s=0.5 0.2741 b8/s0.5 0.3617 b8/s0.5 -0.0876 0.0016 sit-xl-2-webssl-vit-dino1b_full2b_224-irepa conclusive
REPA vs iREPA (webssl-vit-dino1b_full2b_224) imagenet_cls top1 block8@s=0.7 0.0818 b8/s0.7 0.1003 b8/s0.7 -0.0184 0.0024 sit-xl-2-webssl-vit-dino1b_full2b_224-irepa conclusive
REPA vs iREPA (webssl-vit-dino1b_full2b_224) nyu_depth rmse best(layer,s) 0.6790 b16/s0.1 0.6044 b12/s0.1 0.0746 0.0042 sit-xl-2-webssl-vit-dino1b_full2b_224-irepa conclusive
REPA vs iREPA (webssl-vit-dino1b_full2b_224) nyu_depth rmse block8@s=0.1 0.7024 b8/s0.1 0.6309 b8/s0.1 0.0715 0.0088 sit-xl-2-webssl-vit-dino1b_full2b_224-irepa conclusive
REPA vs iREPA (webssl-vit-dino1b_full2b_224) nyu_depth rmse block8@s=0.3 0.7339 b8/s0.3 0.6606 b8/s0.3 0.0733 0.0047 sit-xl-2-webssl-vit-dino1b_full2b_224-irepa conclusive
REPA vs iREPA (webssl-vit-dino1b_full2b_224) nyu_depth rmse block8@s=0.5 0.8486 b8/s0.5 0.7989 b8/s0.5 0.0496 0.0051 sit-xl-2-webssl-vit-dino1b_full2b_224-irepa conclusive
REPA vs iREPA (webssl-vit-dino1b_full2b_224) nyu_depth rmse block8@s=0.7 1.0077 b8/s0.7 0.9963 b8/s0.7 0.0114 0.0226 sit-xl-2-webssl-vit-dino1b_full2b_224-irepa inconclusive (<2 sd)
REPA vs iREPA (webssl-vit-dino1b_full2b_224) nyu_depth delta1 best(layer,s) 0.6634 b16/s0.1 0.7264 b12/s0.1 -0.0630 0.0031 sit-xl-2-webssl-vit-dino1b_full2b_224-irepa conclusive
REPA vs iREPA (webssl-vit-dino1b_full2b_224) nyu_depth delta1 block8@s=0.1 0.6454 b8/s0.1 0.7081 b8/s0.1 -0.0626 0.0054 sit-xl-2-webssl-vit-dino1b_full2b_224-irepa conclusive
REPA vs iREPA (webssl-vit-dino1b_full2b_224) nyu_depth delta1 block8@s=0.3 0.6204 b8/s0.3 0.6834 b8/s0.3 -0.0630 0.0038 sit-xl-2-webssl-vit-dino1b_full2b_224-irepa conclusive
REPA vs iREPA (webssl-vit-dino1b_full2b_224) nyu_depth delta1 block8@s=0.5 0.5479 b8/s0.5 0.5870 b8/s0.5 -0.0391 0.0051 sit-xl-2-webssl-vit-dino1b_full2b_224-irepa conclusive
REPA vs iREPA (webssl-vit-dino1b_full2b_224) nyu_depth delta1 block8@s=0.7 0.4678 b8/s0.7 0.4729 b8/s0.7 -0.0051 0.0094 sit-xl-2-webssl-vit-dino1b_full2b_224-irepa inconclusive (<2 sd)
REPA vs iREPA (webssl-vit-dino1b_full2b_224) latent_recon r2 best(layer,s) 0.9840 b28/s0.1 0.9839 b24/s0.1 0.0000 0.0000 sit-xl-2-webssl-vit-dino1b_full2b_224-repa conclusive
REPA vs iREPA (webssl-vit-dino1b_full2b_224) latent_recon r2 block8@s=0.1 0.9823 b8/s0.1 0.9801 b8/s0.1 0.0022 0.0000 sit-xl-2-webssl-vit-dino1b_full2b_224-repa conclusive
REPA vs iREPA (webssl-vit-dino1b_full2b_224) latent_recon r2 block8@s=0.3 0.8491 b8/s0.3 0.8361 b8/s0.3 0.0130 0.0000 sit-xl-2-webssl-vit-dino1b_full2b_224-repa conclusive
REPA vs iREPA (webssl-vit-dino1b_full2b_224) latent_recon r2 block8@s=0.5 0.6510 b8/s0.5 0.6351 b8/s0.5 0.0160 0.0000 sit-xl-2-webssl-vit-dino1b_full2b_224-repa conclusive
REPA vs iREPA (webssl-vit-dino1b_full2b_224) latent_recon r2 block8@s=0.7 0.4584 b8/s0.7 0.4522 b8/s0.7 0.0063 0.0000 sit-xl-2-webssl-vit-dino1b_full2b_224-repa conclusive
REPA vs iREPA (webssl-vit-dino1b_full2b_224) cknna cknna_teacher best(layer,s) 0.2687 b8/s0.1 0.4789 b8/s0.1 -0.2102 0.0000 sit-xl-2-webssl-vit-dino1b_full2b_224-irepa n/a (no seed std)
REPA vs iREPA (webssl-vit-dino1b_full2b_224) cknna cknna_teacher block8@s=0.1 0.2687 b8/s0.1 0.4789 b8/s0.1 -0.2102 0.0000 sit-xl-2-webssl-vit-dino1b_full2b_224-irepa n/a (no seed std)
REPA vs iREPA (webssl-vit-dino1b_full2b_224) cknna cknna_teacher block8@s=0.3 0.2454 b8/s0.3 0.4244 b8/s0.3 -0.1790 0.0000 sit-xl-2-webssl-vit-dino1b_full2b_224-irepa n/a (no seed std)
REPA vs iREPA (webssl-vit-dino1b_full2b_224) cknna cknna_teacher block8@s=0.5 0.1503 b8/s0.5 0.2493 b8/s0.5 -0.0991 0.0000 sit-xl-2-webssl-vit-dino1b_full2b_224-irepa n/a (no seed std)
REPA vs iREPA (webssl-vit-dino1b_full2b_224) cknna cknna_teacher block8@s=0.7 0.0375 b8/s0.7 0.0485 b8/s0.7 -0.0109 0.0000 sit-xl-2-webssl-vit-dino1b_full2b_224-irepa n/a (no seed std)
REPA vs iREPA (webssl-vit-dino1b_full2b_224) rank pr_token best(layer,s) 113.4706 b8/s0.3 180.6321 b8/s0.1 -67.1615 0.0000 n/a (no seed std)
REPA vs iREPA (webssl-vit-dino1b_full2b_224) rank pr_token block8@s=0.1 97.5150 b8/s0.1 180.6321 b8/s0.1 -83.1171 0.0000 n/a (no seed std)
REPA vs iREPA (webssl-vit-dino1b_full2b_224) rank pr_token block8@s=0.3 113.4706 b8/s0.3 171.5787 b8/s0.3 -58.1081 0.0000 n/a (no seed std)
REPA vs iREPA (webssl-vit-dino1b_full2b_224) rank pr_token block8@s=0.5 106.8010 b8/s0.5 136.8505 b8/s0.5 -30.0496 0.0000 n/a (no seed std)
REPA vs iREPA (webssl-vit-dino1b_full2b_224) rank pr_token block8@s=0.7 76.5038 b8/s0.7 74.9019 b8/s0.7 1.6019 0.0000 n/a (no seed std)
REPA vs iREPA (pe-vit-g) imagenet_cls top1 best(layer,s) 0.4576 b8/s0.1 0.5668 b8/s0.1 -0.1092 0.0065 sit-xl-2-pe-vit-g-irepa conclusive
REPA vs iREPA (pe-vit-g) imagenet_cls top1 block8@s=0.1 0.4576 b8/s0.1 0.5668 b8/s0.1 -0.1092 0.0065 sit-xl-2-pe-vit-g-irepa conclusive
REPA vs iREPA (pe-vit-g) imagenet_cls top1 block8@s=0.3 0.4106 b8/s0.3 0.5188 b8/s0.3 -0.1082 0.0026 sit-xl-2-pe-vit-g-irepa conclusive
REPA vs iREPA (pe-vit-g) imagenet_cls top1 block8@s=0.5 0.2696 b8/s0.5 0.3416 b8/s0.5 -0.0720 0.0022 sit-xl-2-pe-vit-g-irepa conclusive
REPA vs iREPA (pe-vit-g) imagenet_cls top1 block8@s=0.7 0.0872 b8/s0.7 0.0963 b8/s0.7 -0.0091 0.0015 sit-xl-2-pe-vit-g-irepa conclusive
REPA vs iREPA (pe-vit-g) nyu_depth rmse best(layer,s) 0.6937 b16/s0.1 0.6379 b12/s0.1 0.0559 0.0023 sit-xl-2-pe-vit-g-irepa conclusive
REPA vs iREPA (pe-vit-g) nyu_depth rmse block8@s=0.1 0.7352 b8/s0.1 0.6883 b8/s0.1 0.0469 0.0101 sit-xl-2-pe-vit-g-irepa conclusive
REPA vs iREPA (pe-vit-g) nyu_depth rmse block8@s=0.3 0.7684 b8/s0.3 0.7176 b8/s0.3 0.0508 0.0027 sit-xl-2-pe-vit-g-irepa conclusive
REPA vs iREPA (pe-vit-g) nyu_depth rmse block8@s=0.5 0.8606 b8/s0.5 0.8415 b8/s0.5 0.0191 0.0032 sit-xl-2-pe-vit-g-irepa conclusive
REPA vs iREPA (pe-vit-g) nyu_depth rmse block8@s=0.7 0.9970 b8/s0.7 1.0134 b8/s0.7 -0.0164 0.0165 sit-xl-2-pe-vit-g-repa inconclusive (<2 sd)
REPA vs iREPA (pe-vit-g) nyu_depth delta1 best(layer,s) 0.6499 b16/s0.1 0.6976 b12/s0.1 -0.0478 0.0011 sit-xl-2-pe-vit-g-irepa conclusive
REPA vs iREPA (pe-vit-g) nyu_depth delta1 block8@s=0.1 0.6154 b8/s0.1 0.6556 b8/s0.1 -0.0402 0.0072 sit-xl-2-pe-vit-g-irepa conclusive
REPA vs iREPA (pe-vit-g) nyu_depth delta1 block8@s=0.3 0.5898 b8/s0.3 0.6308 b8/s0.3 -0.0411 0.0013 sit-xl-2-pe-vit-g-irepa conclusive
REPA vs iREPA (pe-vit-g) nyu_depth delta1 block8@s=0.5 0.5413 b8/s0.5 0.5533 b8/s0.5 -0.0120 0.0029 sit-xl-2-pe-vit-g-irepa conclusive
REPA vs iREPA (pe-vit-g) nyu_depth delta1 block8@s=0.7 0.4712 b8/s0.7 0.4632 b8/s0.7 0.0081 0.0082 sit-xl-2-pe-vit-g-repa inconclusive (<2 sd)
REPA vs iREPA (pe-vit-g) latent_recon r2 best(layer,s) 0.9838 b24/s0.1 0.9833 b24/s0.1 0.0004 0.0000 sit-xl-2-pe-vit-g-repa conclusive
REPA vs iREPA (pe-vit-g) latent_recon r2 block8@s=0.1 0.9824 b8/s0.1 0.9797 b8/s0.1 0.0028 0.0000 sit-xl-2-pe-vit-g-repa conclusive
REPA vs iREPA (pe-vit-g) latent_recon r2 block8@s=0.3 0.8527 b8/s0.3 0.8401 b8/s0.3 0.0126 0.0000 sit-xl-2-pe-vit-g-repa conclusive
REPA vs iREPA (pe-vit-g) latent_recon r2 block8@s=0.5 0.6543 b8/s0.5 0.6416 b8/s0.5 0.0127 0.0000 sit-xl-2-pe-vit-g-repa conclusive
REPA vs iREPA (pe-vit-g) latent_recon r2 block8@s=0.7 0.4591 b8/s0.7 0.4556 b8/s0.7 0.0035 0.0000 sit-xl-2-pe-vit-g-repa conclusive
REPA vs iREPA (pe-vit-g) cknna cknna_teacher best(layer,s) 0.2063 b8/s0.1 0.3407 b8/s0.1 -0.1345 0.0000 sit-xl-2-pe-vit-g-irepa n/a (no seed std)
REPA vs iREPA (pe-vit-g) cknna cknna_teacher block8@s=0.1 0.2063 b8/s0.1 0.3407 b8/s0.1 -0.1345 0.0000 sit-xl-2-pe-vit-g-irepa n/a (no seed std)
REPA vs iREPA (pe-vit-g) cknna cknna_teacher block8@s=0.3 0.1863 b8/s0.3 0.2989 b8/s0.3 -0.1125 0.0000 sit-xl-2-pe-vit-g-irepa n/a (no seed std)
REPA vs iREPA (pe-vit-g) cknna cknna_teacher block8@s=0.5 0.1075 b8/s0.5 0.1771 b8/s0.5 -0.0696 0.0000 sit-xl-2-pe-vit-g-irepa n/a (no seed std)
REPA vs iREPA (pe-vit-g) cknna cknna_teacher block8@s=0.7 0.0316 b8/s0.7 0.0393 b8/s0.7 -0.0078 0.0000 sit-xl-2-pe-vit-g-irepa n/a (no seed std)
REPA vs iREPA (pe-vit-g) rank pr_token best(layer,s) 90.6914 b28/s0.1 178.9583 b12/s0.5 -88.2668 0.0000 n/a (no seed std)
REPA vs iREPA (pe-vit-g) rank pr_token block8@s=0.1 69.9850 b8/s0.1 113.5463 b8/s0.1 -43.5614 0.0000 n/a (no seed std)
REPA vs iREPA (pe-vit-g) rank pr_token block8@s=0.3 85.3089 b8/s0.3 113.6546 b8/s0.3 -28.3457 0.0000 n/a (no seed std)
REPA vs iREPA (pe-vit-g) rank pr_token block8@s=0.5 82.6109 b8/s0.5 91.6839 b8/s0.5 -9.0730 0.0000 n/a (no seed std)
REPA vs iREPA (pe-vit-g) rank pr_token block8@s=0.7 60.5237 b8/s0.7 47.4260 b8/s0.7 13.0978 0.0000 n/a (no seed std)
REPA vs iREPA (langpe-vit-g) imagenet_cls top1 best(layer,s) 0.4543 b8/s0.1 0.5994 b8/s0.1 -0.1451 0.0014 sit-xl-2-langpe-vit-g-irepa conclusive
REPA vs iREPA (langpe-vit-g) imagenet_cls top1 block8@s=0.1 0.4543 b8/s0.1 0.5994 b8/s0.1 -0.1451 0.0014 sit-xl-2-langpe-vit-g-irepa conclusive
REPA vs iREPA (langpe-vit-g) imagenet_cls top1 block8@s=0.3 0.4093 b8/s0.3 0.5446 b8/s0.3 -0.1353 0.0015 sit-xl-2-langpe-vit-g-irepa conclusive
REPA vs iREPA (langpe-vit-g) imagenet_cls top1 block8@s=0.5 0.2731 b8/s0.5 0.3588 b8/s0.5 -0.0856 0.0024 sit-xl-2-langpe-vit-g-irepa conclusive
REPA vs iREPA (langpe-vit-g) imagenet_cls top1 block8@s=0.7 0.0837 b8/s0.7 0.1015 b8/s0.7 -0.0178 0.0007 sit-xl-2-langpe-vit-g-irepa conclusive
REPA vs iREPA (langpe-vit-g) nyu_depth rmse best(layer,s) 0.6683 b16/s0.1 0.6072 b12/s0.1 0.0610 0.0050 sit-xl-2-langpe-vit-g-irepa conclusive
REPA vs iREPA (langpe-vit-g) nyu_depth rmse block8@s=0.1 0.6947 b8/s0.1 0.6413 b8/s0.1 0.0535 0.0034 sit-xl-2-langpe-vit-g-irepa conclusive
REPA vs iREPA (langpe-vit-g) nyu_depth rmse block8@s=0.3 0.7283 b8/s0.3 0.6722 b8/s0.3 0.0561 0.0058 sit-xl-2-langpe-vit-g-irepa conclusive
REPA vs iREPA (langpe-vit-g) nyu_depth rmse block8@s=0.5 0.8366 b8/s0.5 0.7925 b8/s0.5 0.0441 0.0031 sit-xl-2-langpe-vit-g-irepa conclusive
REPA vs iREPA (langpe-vit-g) nyu_depth rmse block8@s=0.7 0.9902 b8/s0.7 0.9958 b8/s0.7 -0.0057 0.0022 sit-xl-2-langpe-vit-g-repa conclusive
REPA vs iREPA (langpe-vit-g) nyu_depth delta1 best(layer,s) 0.6729 b16/s0.1 0.7218 b12/s0.1 -0.0489 0.0031 sit-xl-2-langpe-vit-g-irepa conclusive
REPA vs iREPA (langpe-vit-g) nyu_depth delta1 block8@s=0.1 0.6481 b8/s0.1 0.6947 b8/s0.1 -0.0465 0.0014 sit-xl-2-langpe-vit-g-irepa conclusive
REPA vs iREPA (langpe-vit-g) nyu_depth delta1 block8@s=0.3 0.6251 b8/s0.3 0.6696 b8/s0.3 -0.0446 0.0038 sit-xl-2-langpe-vit-g-irepa conclusive
REPA vs iREPA (langpe-vit-g) nyu_depth delta1 block8@s=0.5 0.5566 b8/s0.5 0.5836 b8/s0.5 -0.0270 0.0040 sit-xl-2-langpe-vit-g-irepa conclusive
REPA vs iREPA (langpe-vit-g) nyu_depth delta1 block8@s=0.7 0.4710 b8/s0.7 0.4740 b8/s0.7 -0.0030 0.0012 sit-xl-2-langpe-vit-g-irepa conclusive
REPA vs iREPA (langpe-vit-g) latent_recon r2 best(layer,s) 0.9839 b28/s0.1 0.9833 b24/s0.1 0.0006 0.0000 sit-xl-2-langpe-vit-g-repa conclusive
REPA vs iREPA (langpe-vit-g) latent_recon r2 block8@s=0.1 0.9826 b8/s0.1 0.9789 b8/s0.1 0.0037 0.0000 sit-xl-2-langpe-vit-g-repa conclusive
REPA vs iREPA (langpe-vit-g) latent_recon r2 block8@s=0.3 0.8546 b8/s0.3 0.8351 b8/s0.3 0.0195 0.0000 sit-xl-2-langpe-vit-g-repa conclusive
REPA vs iREPA (langpe-vit-g) latent_recon r2 block8@s=0.5 0.6580 b8/s0.5 0.6365 b8/s0.5 0.0216 0.0000 sit-xl-2-langpe-vit-g-repa conclusive
REPA vs iREPA (langpe-vit-g) latent_recon r2 block8@s=0.7 0.4616 b8/s0.7 0.4546 b8/s0.7 0.0070 0.0000 sit-xl-2-langpe-vit-g-repa conclusive
REPA vs iREPA (langpe-vit-g) cknna cknna_teacher best(layer,s) 0.2521 b8/s0.1 0.5022 b8/s0.1 -0.2501 0.0000 sit-xl-2-langpe-vit-g-irepa n/a (no seed std)
REPA vs iREPA (langpe-vit-g) cknna cknna_teacher block8@s=0.1 0.2521 b8/s0.1 0.5022 b8/s0.1 -0.2501 0.0000 sit-xl-2-langpe-vit-g-irepa n/a (no seed std)
REPA vs iREPA (langpe-vit-g) cknna cknna_teacher block8@s=0.3 0.2350 b8/s0.3 0.4440 b8/s0.3 -0.2090 0.0000 sit-xl-2-langpe-vit-g-irepa n/a (no seed std)
REPA vs iREPA (langpe-vit-g) cknna cknna_teacher block8@s=0.5 0.1456 b8/s0.5 0.2646 b8/s0.5 -0.1190 0.0000 sit-xl-2-langpe-vit-g-irepa n/a (no seed std)
REPA vs iREPA (langpe-vit-g) cknna cknna_teacher block8@s=0.7 0.0375 b8/s0.7 0.0540 b8/s0.7 -0.0166 0.0000 sit-xl-2-langpe-vit-g-irepa n/a (no seed std)
REPA vs iREPA (langpe-vit-g) rank pr_token best(layer,s) 87.2289 b8/s0.3 262.7118 b8/s0.1 -175.4830 0.0000 n/a (no seed std)
REPA vs iREPA (langpe-vit-g) rank pr_token block8@s=0.1 71.0472 b8/s0.1 262.7118 b8/s0.1 -191.6646 0.0000 n/a (no seed std)
REPA vs iREPA (langpe-vit-g) rank pr_token block8@s=0.3 87.2289 b8/s0.3 260.8706 b8/s0.3 -173.6417 0.0000 n/a (no seed std)
REPA vs iREPA (langpe-vit-g) rank pr_token block8@s=0.5 86.1099 b8/s0.5 208.0252 b8/s0.5 -121.9153 0.0000 n/a (no seed std)
REPA vs iREPA (langpe-vit-g) rank pr_token block8@s=0.7 63.9491 b8/s0.7 105.7177 b8/s0.7 -41.7687 0.0000 n/a (no seed std)
REPA-dinov2 400K vs SiT 400K imagenet_cls top1 best(layer,s) 0.5492 b8/s0.1 0.2865 b20/s0.1 0.2627 0.0025 sit-xl-2-dinov2-vit-b-repa conclusive
REPA-dinov2 400K vs SiT 400K imagenet_cls top1 block8@s=0.1 0.5492 b8/s0.1 0.0447 b8/s0.1 0.5045 0.0026 sit-xl-2-dinov2-vit-b-repa conclusive
REPA-dinov2 400K vs SiT 400K imagenet_cls top1 block8@s=0.3 0.5012 b8/s0.3 0.0389 b8/s0.3 0.4624 0.0022 sit-xl-2-dinov2-vit-b-repa conclusive
REPA-dinov2 400K vs SiT 400K imagenet_cls top1 block8@s=0.5 0.3396 b8/s0.5 0.0307 b8/s0.5 0.3089 0.0007 sit-xl-2-dinov2-vit-b-repa conclusive
REPA-dinov2 400K vs SiT 400K imagenet_cls top1 block8@s=0.7 0.1076 b8/s0.7 0.0192 b8/s0.7 0.0884 0.0016 sit-xl-2-dinov2-vit-b-repa conclusive
REPA-dinov2 400K vs SiT 400K nyu_depth rmse best(layer,s) 0.6236 b12/s0.1 0.8096 b20/s0.3 -0.1859 0.0079 sit-xl-2-dinov2-vit-b-repa conclusive
REPA-dinov2 400K vs SiT 400K nyu_depth rmse block8@s=0.1 0.6423 b8/s0.1 1.0076 b8/s0.1 -0.3652 0.0042 sit-xl-2-dinov2-vit-b-repa conclusive
REPA-dinov2 400K vs SiT 400K nyu_depth rmse block8@s=0.3 0.6804 b8/s0.3 1.0057 b8/s0.3 -0.3253 0.0049 sit-xl-2-dinov2-vit-b-repa conclusive
REPA-dinov2 400K vs SiT 400K nyu_depth rmse block8@s=0.5 0.7970 b8/s0.5 1.0133 b8/s0.5 -0.2163 0.0019 sit-xl-2-dinov2-vit-b-repa conclusive
REPA-dinov2 400K vs SiT 400K nyu_depth rmse block8@s=0.7 0.9916 b8/s0.7 1.0370 b8/s0.7 -0.0454 0.0067 sit-xl-2-dinov2-vit-b-repa conclusive
REPA-dinov2 400K vs SiT 400K nyu_depth delta1 best(layer,s) 0.7130 b12/s0.1 0.5749 b20/s0.3 0.1381 0.0053 sit-xl-2-dinov2-vit-b-repa conclusive
REPA-dinov2 400K vs SiT 400K nyu_depth delta1 block8@s=0.1 0.6967 b8/s0.1 0.4589 b8/s0.1 0.2378 0.0010 sit-xl-2-dinov2-vit-b-repa conclusive
REPA-dinov2 400K vs SiT 400K nyu_depth delta1 block8@s=0.3 0.6666 b8/s0.3 0.4568 b8/s0.3 0.2098 0.0014 sit-xl-2-dinov2-vit-b-repa conclusive
REPA-dinov2 400K vs SiT 400K nyu_depth delta1 block8@s=0.5 0.5854 b8/s0.5 0.4534 b8/s0.5 0.1321 0.0014 sit-xl-2-dinov2-vit-b-repa conclusive
REPA-dinov2 400K vs SiT 400K nyu_depth delta1 block8@s=0.7 0.4742 b8/s0.7 0.4467 b8/s0.7 0.0275 0.0044 sit-xl-2-dinov2-vit-b-repa conclusive
REPA-dinov2 400K vs SiT 400K latent_recon r2 best(layer,s) 0.9842 b24/s0.1 0.9845 b28/s0.1 -0.0003 0.0000 sit-baseline-400k conclusive
REPA-dinov2 400K vs SiT 400K latent_recon r2 block8@s=0.1 0.9812 b8/s0.1 0.9816 b8/s0.1 -0.0003 0.0000 sit-baseline-400k conclusive
REPA-dinov2 400K vs SiT 400K latent_recon r2 block8@s=0.3 0.8453 b8/s0.3 0.8310 b8/s0.3 0.0143 0.0000 sit-xl-2-dinov2-vit-b-repa conclusive
REPA-dinov2 400K vs SiT 400K latent_recon r2 block8@s=0.5 0.6452 b8/s0.5 0.6103 b8/s0.5 0.0349 0.0000 sit-xl-2-dinov2-vit-b-repa conclusive
REPA-dinov2 400K vs SiT 400K latent_recon r2 block8@s=0.7 0.4565 b8/s0.7 0.4177 b8/s0.7 0.0389 0.0000 sit-xl-2-dinov2-vit-b-repa conclusive
REPA-dinov2 400K vs SiT 400K cknna cknna_dinov2 best(layer,s) 0.4955 b8/s0.1 0.0933 b24/s0.3 0.4023 0.0000 sit-xl-2-dinov2-vit-b-repa n/a (no seed std)
REPA-dinov2 400K vs SiT 400K cknna cknna_dinov2 block8@s=0.1 0.4955 b8/s0.1 0.0118 b8/s0.1 0.4837 0.0000 sit-xl-2-dinov2-vit-b-repa n/a (no seed std)
REPA-dinov2 400K vs SiT 400K cknna cknna_dinov2 block8@s=0.3 0.4522 b8/s0.3 0.0107 b8/s0.3 0.4415 0.0000 sit-xl-2-dinov2-vit-b-repa n/a (no seed std)
REPA-dinov2 400K vs SiT 400K cknna cknna_dinov2 block8@s=0.5 0.2882 b8/s0.5 0.0099 b8/s0.5 0.2783 0.0000 sit-xl-2-dinov2-vit-b-repa n/a (no seed std)
REPA-dinov2 400K vs SiT 400K cknna cknna_dinov2 block8@s=0.7 0.0697 b8/s0.7 0.0076 b8/s0.7 0.0621 0.0000 sit-xl-2-dinov2-vit-b-repa n/a (no seed std)
REPA-dinov2 400K vs SiT 400K rank pr_token best(layer,s) 208.8756 b8/s0.3 145.0956 b28/s0.1 63.7800 0.0000 n/a (no seed std)
REPA-dinov2 400K vs SiT 400K rank pr_token block8@s=0.1 190.6034 b8/s0.1 17.0594 b8/s0.1 173.5441 0.0000 n/a (no seed std)
REPA-dinov2 400K vs SiT 400K rank pr_token block8@s=0.3 208.8756 b8/s0.3 19.3032 b8/s0.3 189.5725 0.0000 n/a (no seed std)
REPA-dinov2 400K vs SiT 400K rank pr_token block8@s=0.5 184.5681 b8/s0.5 22.3664 b8/s0.5 162.2017 0.0000 n/a (no seed std)
REPA-dinov2 400K vs SiT 400K rank pr_token block8@s=0.7 113.6005 b8/s0.7 23.4953 b8/s0.7 90.1052 0.0000 n/a (no seed std)
iREPA-dinov2 400K vs SiT 400K imagenet_cls top1 best(layer,s) 0.5827 b8/s0.1 0.2865 b20/s0.1 0.2962 0.0024 sit-xl-2-dinov2-vit-b-irepa conclusive
iREPA-dinov2 400K vs SiT 400K imagenet_cls top1 block8@s=0.1 0.5827 b8/s0.1 0.0447 b8/s0.1 0.5380 0.0025 sit-xl-2-dinov2-vit-b-irepa conclusive
iREPA-dinov2 400K vs SiT 400K imagenet_cls top1 block8@s=0.3 0.5377 b8/s0.3 0.0389 b8/s0.3 0.4988 0.0016 sit-xl-2-dinov2-vit-b-irepa conclusive
iREPA-dinov2 400K vs SiT 400K imagenet_cls top1 block8@s=0.5 0.3641 b8/s0.5 0.0307 b8/s0.5 0.3334 0.0009 sit-xl-2-dinov2-vit-b-irepa conclusive
iREPA-dinov2 400K vs SiT 400K imagenet_cls top1 block8@s=0.7 0.1042 b8/s0.7 0.0192 b8/s0.7 0.0850 0.0013 sit-xl-2-dinov2-vit-b-irepa conclusive
iREPA-dinov2 400K vs SiT 400K nyu_depth rmse best(layer,s) 0.5993 b8/s0.1 0.8096 b20/s0.3 -0.2103 0.0032 sit-xl-2-dinov2-vit-b-irepa conclusive
iREPA-dinov2 400K vs SiT 400K nyu_depth rmse block8@s=0.1 0.5993 b8/s0.1 1.0076 b8/s0.1 -0.4083 0.0045 sit-xl-2-dinov2-vit-b-irepa conclusive
iREPA-dinov2 400K vs SiT 400K nyu_depth rmse block8@s=0.3 0.6378 b8/s0.3 1.0057 b8/s0.3 -0.3679 0.0049 sit-xl-2-dinov2-vit-b-irepa conclusive
iREPA-dinov2 400K vs SiT 400K nyu_depth rmse block8@s=0.5 0.7863 b8/s0.5 1.0133 b8/s0.5 -0.2271 0.0076 sit-xl-2-dinov2-vit-b-irepa conclusive
iREPA-dinov2 400K vs SiT 400K nyu_depth rmse block8@s=0.7 0.9946 b8/s0.7 1.0370 b8/s0.7 -0.0424 0.0064 sit-xl-2-dinov2-vit-b-irepa conclusive
iREPA-dinov2 400K vs SiT 400K nyu_depth delta1 best(layer,s) 0.7342 b8/s0.1 0.5749 b20/s0.3 0.1593 0.0018 sit-xl-2-dinov2-vit-b-irepa conclusive
iREPA-dinov2 400K vs SiT 400K nyu_depth delta1 block8@s=0.1 0.7342 b8/s0.1 0.4589 b8/s0.1 0.2754 0.0009 sit-xl-2-dinov2-vit-b-irepa conclusive
iREPA-dinov2 400K vs SiT 400K nyu_depth delta1 block8@s=0.3 0.7017 b8/s0.3 0.4568 b8/s0.3 0.2449 0.0012 sit-xl-2-dinov2-vit-b-irepa conclusive
iREPA-dinov2 400K vs SiT 400K nyu_depth delta1 block8@s=0.5 0.5945 b8/s0.5 0.4534 b8/s0.5 0.1411 0.0048 sit-xl-2-dinov2-vit-b-irepa conclusive
iREPA-dinov2 400K vs SiT 400K nyu_depth delta1 block8@s=0.7 0.4738 b8/s0.7 0.4467 b8/s0.7 0.0271 0.0044 sit-xl-2-dinov2-vit-b-irepa conclusive
iREPA-dinov2 400K vs SiT 400K latent_recon r2 best(layer,s) 0.9843 b28/s0.1 0.9845 b28/s0.1 -0.0001 0.0000 sit-baseline-400k conclusive
iREPA-dinov2 400K vs SiT 400K latent_recon r2 block8@s=0.1 0.9799 b8/s0.1 0.9816 b8/s0.1 -0.0017 0.0000 sit-baseline-400k conclusive
iREPA-dinov2 400K vs SiT 400K latent_recon r2 block8@s=0.3 0.8364 b8/s0.3 0.8310 b8/s0.3 0.0054 0.0000 sit-xl-2-dinov2-vit-b-irepa conclusive
iREPA-dinov2 400K vs SiT 400K latent_recon r2 block8@s=0.5 0.6348 b8/s0.5 0.6103 b8/s0.5 0.0245 0.0000 sit-xl-2-dinov2-vit-b-irepa conclusive
iREPA-dinov2 400K vs SiT 400K latent_recon r2 block8@s=0.7 0.4520 b8/s0.7 0.4177 b8/s0.7 0.0343 0.0000 sit-xl-2-dinov2-vit-b-irepa conclusive
iREPA-dinov2 400K vs SiT 400K cknna cknna_dinov2 best(layer,s) 0.6284 b8/s0.1 0.0933 b24/s0.3 0.5351 0.0000 sit-xl-2-dinov2-vit-b-irepa n/a (no seed std)
iREPA-dinov2 400K vs SiT 400K cknna cknna_dinov2 block8@s=0.1 0.6284 b8/s0.1 0.0118 b8/s0.1 0.6165 0.0000 sit-xl-2-dinov2-vit-b-irepa n/a (no seed std)
iREPA-dinov2 400K vs SiT 400K cknna cknna_dinov2 block8@s=0.3 0.5602 b8/s0.3 0.0107 b8/s0.3 0.5495 0.0000 sit-xl-2-dinov2-vit-b-irepa n/a (no seed std)
iREPA-dinov2 400K vs SiT 400K cknna cknna_dinov2 block8@s=0.5 0.3408 b8/s0.5 0.0099 b8/s0.5 0.3309 0.0000 sit-xl-2-dinov2-vit-b-irepa n/a (no seed std)
iREPA-dinov2 400K vs SiT 400K cknna cknna_dinov2 block8@s=0.7 0.0745 b8/s0.7 0.0076 b8/s0.7 0.0668 0.0000 sit-xl-2-dinov2-vit-b-irepa n/a (no seed std)
iREPA-dinov2 400K vs SiT 400K rank pr_token best(layer,s) 373.5329 b8/s0.1 145.0956 b28/s0.1 228.4372 0.0000 n/a (no seed std)
iREPA-dinov2 400K vs SiT 400K rank pr_token block8@s=0.1 373.5329 b8/s0.1 17.0594 b8/s0.1 356.4735 0.0000 n/a (no seed std)
iREPA-dinov2 400K vs SiT 400K rank pr_token block8@s=0.3 369.6074 b8/s0.3 19.3032 b8/s0.3 350.3043 0.0000 n/a (no seed std)
iREPA-dinov2 400K vs SiT 400K rank pr_token block8@s=0.5 308.7439 b8/s0.5 22.3664 b8/s0.5 286.3775 0.0000 n/a (no seed std)
iREPA-dinov2 400K vs SiT 400K rank pr_token block8@s=0.7 168.5247 b8/s0.7 23.4953 b8/s0.7 145.0294 0.0000 n/a (no seed std)
REPA 4M vs HASTE imagenet_cls top1 best(layer,s) 0.6389 b8/s0.1 0.5561 b8/s0.1 0.0828 0.0017 repa-4m conclusive
REPA 4M vs HASTE imagenet_cls top1 block8@s=0.1 0.6389 b8/s0.1 0.5561 b8/s0.1 0.0828 0.0017 repa-4m conclusive
REPA 4M vs HASTE imagenet_cls top1 block8@s=0.3 0.6006 b8/s0.3 0.5286 b8/s0.3 0.0720 0.0020 repa-4m conclusive
REPA 4M vs HASTE imagenet_cls top1 block8@s=0.5 0.4334 b8/s0.5 0.3824 b8/s0.5 0.0511 0.0032 repa-4m conclusive
REPA 4M vs HASTE imagenet_cls top1 block8@s=0.7 0.1413 b8/s0.7 0.1164 b8/s0.7 0.0249 0.0050 repa-4m conclusive
REPA 4M vs HASTE nyu_depth rmse best(layer,s) 0.5422 b12/s0.1 0.5742 b12/s0.1 -0.0320 0.0022 repa-4m conclusive
REPA 4M vs HASTE nyu_depth rmse block8@s=0.1 0.5693 b8/s0.1 0.5750 b8/s0.1 -0.0057 0.0018 repa-4m conclusive
REPA 4M vs HASTE nyu_depth rmse block8@s=0.3 0.6075 b8/s0.3 0.6080 b8/s0.3 -0.0005 0.0022 repa-4m inconclusive (<2 sd)
REPA 4M vs HASTE nyu_depth rmse block8@s=0.5 0.7603 b8/s0.5 0.7626 b8/s0.5 -0.0022 0.0078 repa-4m inconclusive (<2 sd)
REPA 4M vs HASTE nyu_depth rmse block8@s=0.7 1.0116 b8/s0.7 0.9932 b8/s0.7 0.0184 0.0279 haste inconclusive (<2 sd)
REPA 4M vs HASTE nyu_depth delta1 best(layer,s) 0.7775 b12/s0.1 0.7545 b8/s0.1 0.0230 0.0007 repa-4m conclusive
REPA 4M vs HASTE nyu_depth delta1 block8@s=0.1 0.7566 b8/s0.1 0.7545 b8/s0.1 0.0021 0.0011 repa-4m inconclusive (<2 sd)
REPA 4M vs HASTE nyu_depth delta1 block8@s=0.3 0.7241 b8/s0.3 0.7283 b8/s0.3 -0.0042 0.0015 haste conclusive
REPA 4M vs HASTE nyu_depth delta1 block8@s=0.5 0.6077 b8/s0.5 0.6113 b8/s0.5 -0.0036 0.0069 haste inconclusive (<2 sd)
REPA 4M vs HASTE nyu_depth delta1 block8@s=0.7 0.4594 b8/s0.7 0.4718 b8/s0.7 -0.0125 0.0165 haste inconclusive (<2 sd)
REPA 4M vs HASTE latent_recon r2 best(layer,s) 0.9841 b24/s0.1 0.9843 b24/s0.1 -0.0003 0.0000 haste conclusive
REPA 4M vs HASTE latent_recon r2 block8@s=0.1 0.9800 b8/s0.1 0.9822 b8/s0.1 -0.0022 0.0000 haste conclusive
REPA 4M vs HASTE latent_recon r2 block8@s=0.3 0.8360 b8/s0.3 0.8443 b8/s0.3 -0.0083 0.0000 haste conclusive
REPA 4M vs HASTE latent_recon r2 block8@s=0.5 0.6373 b8/s0.5 0.6428 b8/s0.5 -0.0055 0.0000 haste conclusive
REPA 4M vs HASTE latent_recon r2 block8@s=0.7 0.4556 b8/s0.7 0.4557 b8/s0.7 -0.0001 0.0000 haste conclusive
REPA 4M vs HASTE cknna cknna_dinov2 best(layer,s) 0.5852 b8/s0.1 0.3966 b8/s0.3 0.1886 0.0000 repa-4m n/a (no seed std)
REPA 4M vs HASTE cknna cknna_dinov2 block8@s=0.1 0.5852 b8/s0.1 0.3932 b8/s0.1 0.1920 0.0000 repa-4m n/a (no seed std)
REPA 4M vs HASTE cknna cknna_dinov2 block8@s=0.3 0.5590 b8/s0.3 0.3966 b8/s0.3 0.1623 0.0000 repa-4m n/a (no seed std)
REPA 4M vs HASTE cknna cknna_dinov2 block8@s=0.5 0.3965 b8/s0.5 0.2758 b8/s0.5 0.1207 0.0000 repa-4m n/a (no seed std)
REPA 4M vs HASTE cknna cknna_dinov2 block8@s=0.7 0.0975 b8/s0.7 0.0663 b8/s0.7 0.0312 0.0000 repa-4m n/a (no seed std)
REPA 4M vs HASTE rank pr_token best(layer,s) 245.2866 b8/s0.3 72.4689 b28/s0.3 172.8177 0.0000 n/a (no seed std)
REPA 4M vs HASTE rank pr_token block8@s=0.1 196.3902 b8/s0.1 19.8555 b8/s0.1 176.5347 0.0000 n/a (no seed std)
REPA 4M vs HASTE rank pr_token block8@s=0.3 245.2866 b8/s0.3 17.4419 b8/s0.3 227.8447 0.0000 n/a (no seed std)
REPA 4M vs HASTE rank pr_token block8@s=0.5 238.6067 b8/s0.5 15.0187 b8/s0.5 223.5880 0.0000 n/a (no seed std)
REPA 4M vs HASTE rank pr_token block8@s=0.7 138.9348 b8/s0.7 11.3167 b8/s0.7 127.6181 0.0000 n/a (no seed std)
REPA 400K vs SiT 7M (FID-matched control) imagenet_cls top1 best(layer,s) 0.5492 b8/s0.1 0.3837 b20/s0.1 0.1655 0.0031 sit-xl-2-dinov2-vit-b-repa conclusive
REPA 400K vs SiT 7M (FID-matched control) imagenet_cls top1 block8@s=0.1 0.5492 b8/s0.1 0.1142 b8/s0.1 0.4350 0.0022 sit-xl-2-dinov2-vit-b-repa conclusive
REPA 400K vs SiT 7M (FID-matched control) imagenet_cls top1 block8@s=0.3 0.5012 b8/s0.3 0.0975 b8/s0.3 0.4037 0.0028 sit-xl-2-dinov2-vit-b-repa conclusive
REPA 400K vs SiT 7M (FID-matched control) imagenet_cls top1 block8@s=0.5 0.3396 b8/s0.5 0.0650 b8/s0.5 0.2746 0.0014 sit-xl-2-dinov2-vit-b-repa conclusive
REPA 400K vs SiT 7M (FID-matched control) imagenet_cls top1 block8@s=0.7 0.1076 b8/s0.7 0.0331 b8/s0.7 0.0746 0.0019 sit-xl-2-dinov2-vit-b-repa conclusive
REPA 400K vs SiT 7M (FID-matched control) nyu_depth rmse best(layer,s) 0.6236 b12/s0.1 0.6898 b20/s0.3 -0.0661 0.0074 sit-xl-2-dinov2-vit-b-repa conclusive
REPA 400K vs SiT 7M (FID-matched control) nyu_depth rmse block8@s=0.1 0.6423 b8/s0.1 0.9587 b8/s0.1 -0.3164 0.0012 sit-xl-2-dinov2-vit-b-repa conclusive
REPA 400K vs SiT 7M (FID-matched control) nyu_depth rmse block8@s=0.3 0.6804 b8/s0.3 0.9619 b8/s0.3 -0.2815 0.0041 sit-xl-2-dinov2-vit-b-repa conclusive
REPA 400K vs SiT 7M (FID-matched control) nyu_depth rmse block8@s=0.5 0.7970 b8/s0.5 0.9962 b8/s0.5 -0.1992 0.0068 sit-xl-2-dinov2-vit-b-repa conclusive
REPA 400K vs SiT 7M (FID-matched control) nyu_depth rmse block8@s=0.7 0.9916 b8/s0.7 1.0208 b8/s0.7 -0.0292 0.0021 sit-xl-2-dinov2-vit-b-repa conclusive
REPA 400K vs SiT 7M (FID-matched control) nyu_depth delta1 best(layer,s) 0.7130 b12/s0.1 0.6537 b20/s0.3 0.0593 0.0051 sit-xl-2-dinov2-vit-b-repa conclusive
REPA 400K vs SiT 7M (FID-matched control) nyu_depth delta1 block8@s=0.1 0.6967 b8/s0.1 0.4792 b8/s0.1 0.2175 0.0012 sit-xl-2-dinov2-vit-b-repa conclusive
REPA 400K vs SiT 7M (FID-matched control) nyu_depth delta1 block8@s=0.3 0.6666 b8/s0.3 0.4771 b8/s0.3 0.1894 0.0016 sit-xl-2-dinov2-vit-b-repa conclusive
REPA 400K vs SiT 7M (FID-matched control) nyu_depth delta1 block8@s=0.5 0.5854 b8/s0.5 0.4606 b8/s0.5 0.1248 0.0045 sit-xl-2-dinov2-vit-b-repa conclusive
REPA 400K vs SiT 7M (FID-matched control) nyu_depth delta1 block8@s=0.7 0.4742 b8/s0.7 0.4514 b8/s0.7 0.0227 0.0008 sit-xl-2-dinov2-vit-b-repa conclusive
REPA 400K vs SiT 7M (FID-matched control) latent_recon r2 best(layer,s) 0.9842 b24/s0.1 0.9828 b24/s0.1 0.0014 0.0000 sit-xl-2-dinov2-vit-b-repa conclusive
REPA 400K vs SiT 7M (FID-matched control) latent_recon r2 block8@s=0.1 0.9812 b8/s0.1 0.9816 b8/s0.1 -0.0004 0.0000 sit-7m conclusive
REPA 400K vs SiT 7M (FID-matched control) latent_recon r2 block8@s=0.3 0.8453 b8/s0.3 0.8402 b8/s0.3 0.0051 0.0000 sit-xl-2-dinov2-vit-b-repa conclusive
REPA 400K vs SiT 7M (FID-matched control) latent_recon r2 block8@s=0.5 0.6452 b8/s0.5 0.6333 b8/s0.5 0.0119 0.0000 sit-xl-2-dinov2-vit-b-repa conclusive
REPA 400K vs SiT 7M (FID-matched control) latent_recon r2 block8@s=0.7 0.4565 b8/s0.7 0.4443 b8/s0.7 0.0123 0.0000 sit-xl-2-dinov2-vit-b-repa conclusive
REPA 400K vs SiT 7M (FID-matched control) cknna cknna_dinov2 best(layer,s) 0.4955 b8/s0.1 0.1430 b20/s0.3 0.3525 0.0000 sit-xl-2-dinov2-vit-b-repa n/a (no seed std)
REPA 400K vs SiT 7M (FID-matched control) cknna cknna_dinov2 block8@s=0.1 0.4955 b8/s0.1 0.0143 b8/s0.1 0.4812 0.0000 sit-xl-2-dinov2-vit-b-repa n/a (no seed std)
REPA 400K vs SiT 7M (FID-matched control) cknna cknna_dinov2 block8@s=0.3 0.4522 b8/s0.3 0.0133 b8/s0.3 0.4389 0.0000 sit-xl-2-dinov2-vit-b-repa n/a (no seed std)
REPA 400K vs SiT 7M (FID-matched control) cknna cknna_dinov2 block8@s=0.5 0.2882 b8/s0.5 0.0119 b8/s0.5 0.2763 0.0000 sit-xl-2-dinov2-vit-b-repa n/a (no seed std)
REPA 400K vs SiT 7M (FID-matched control) cknna cknna_dinov2 block8@s=0.7 0.0697 b8/s0.7 0.0086 b8/s0.7 0.0611 0.0000 sit-xl-2-dinov2-vit-b-repa n/a (no seed std)
REPA 400K vs SiT 7M (FID-matched control) rank pr_token best(layer,s) 208.8756 b8/s0.3 35.5541 b28/s0.3 173.3216 0.0000 n/a (no seed std)
REPA 400K vs SiT 7M (FID-matched control) rank pr_token block8@s=0.1 190.6034 b8/s0.1 13.5915 b8/s0.1 177.0119 0.0000 n/a (no seed std)
REPA 400K vs SiT 7M (FID-matched control) rank pr_token block8@s=0.3 208.8756 b8/s0.3 14.6111 b8/s0.3 194.2645 0.0000 n/a (no seed std)
REPA 400K vs SiT 7M (FID-matched control) rank pr_token block8@s=0.5 184.5681 b8/s0.5 17.4588 b8/s0.5 167.1093 0.0000 n/a (no seed std)
REPA 400K vs SiT 7M (FID-matched control) rank pr_token block8@s=0.7 113.6005 b8/s0.7 20.6642 b8/s0.7 92.9363 0.0000 n/a (no seed std)
Dispersive 80ep vs SiT 400K imagenet_cls top1 best(layer,s) 0.3002 b12/s0.3 0.2865 b20/s0.1 0.0137 0.0024 dispersive-80ep conclusive
Dispersive 80ep vs SiT 400K imagenet_cls top1 block8@s=0.1 0.2474 b8/s0.1 0.0447 b8/s0.1 0.2027 0.0018 dispersive-80ep conclusive
Dispersive 80ep vs SiT 400K imagenet_cls top1 block8@s=0.3 0.2433 b8/s0.3 0.0389 b8/s0.3 0.2044 0.0025 dispersive-80ep conclusive
Dispersive 80ep vs SiT 400K imagenet_cls top1 block8@s=0.5 0.1703 b8/s0.5 0.0307 b8/s0.5 0.1396 0.0013 dispersive-80ep conclusive
Dispersive 80ep vs SiT 400K imagenet_cls top1 block8@s=0.7 0.0646 b8/s0.7 0.0192 b8/s0.7 0.0454 0.0009 dispersive-80ep conclusive
Dispersive 80ep vs SiT 400K nyu_depth rmse best(layer,s) 0.7924 b12/s0.3 0.8096 b20/s0.3 -0.0172 0.0032 dispersive-80ep conclusive
Dispersive 80ep vs SiT 400K nyu_depth rmse block8@s=0.1 0.8692 b8/s0.1 1.0076 b8/s0.1 -0.1384 0.0042 dispersive-80ep conclusive
Dispersive 80ep vs SiT 400K nyu_depth rmse block8@s=0.3 0.8737 b8/s0.3 1.0057 b8/s0.3 -0.1320 0.0062 dispersive-80ep conclusive
Dispersive 80ep vs SiT 400K nyu_depth rmse block8@s=0.5 0.9256 b8/s0.5 1.0133 b8/s0.5 -0.0877 0.0084 dispersive-80ep conclusive
Dispersive 80ep vs SiT 400K nyu_depth rmse block8@s=0.7 1.0074 b8/s0.7 1.0370 b8/s0.7 -0.0296 0.0081 dispersive-80ep conclusive
Dispersive 80ep vs SiT 400K nyu_depth delta1 best(layer,s) 0.5835 b12/s0.3 0.5749 b20/s0.3 0.0086 0.0020 dispersive-80ep conclusive
Dispersive 80ep vs SiT 400K nyu_depth delta1 block8@s=0.1 0.5410 b8/s0.1 0.4589 b8/s0.1 0.0821 0.0004 dispersive-80ep conclusive
Dispersive 80ep vs SiT 400K nyu_depth delta1 block8@s=0.3 0.5358 b8/s0.3 0.4568 b8/s0.3 0.0790 0.0033 dispersive-80ep conclusive
Dispersive 80ep vs SiT 400K nyu_depth delta1 block8@s=0.5 0.5080 b8/s0.5 0.4534 b8/s0.5 0.0547 0.0056 dispersive-80ep conclusive
Dispersive 80ep vs SiT 400K nyu_depth delta1 block8@s=0.7 0.4666 b8/s0.7 0.4467 b8/s0.7 0.0199 0.0046 dispersive-80ep conclusive
Dispersive 80ep vs SiT 400K latent_recon r2 best(layer,s) 0.9845 b20/s0.1 0.9845 b28/s0.1 0.0000 0.0000 dispersive-80ep conclusive
Dispersive 80ep vs SiT 400K latent_recon r2 block8@s=0.1 0.9825 b8/s0.1 0.9816 b8/s0.1 0.0010 0.0000 dispersive-80ep conclusive
Dispersive 80ep vs SiT 400K latent_recon r2 block8@s=0.3 0.8618 b8/s0.3 0.8310 b8/s0.3 0.0308 0.0000 dispersive-80ep conclusive
Dispersive 80ep vs SiT 400K latent_recon r2 block8@s=0.5 0.6677 b8/s0.5 0.6103 b8/s0.5 0.0574 0.0000 dispersive-80ep conclusive
Dispersive 80ep vs SiT 400K latent_recon r2 block8@s=0.7 0.4652 b8/s0.7 0.4177 b8/s0.7 0.0476 0.0000 dispersive-80ep conclusive
Dispersive 80ep vs SiT 400K cknna cknna_dinov2 best(layer,s) 0.0976 b12/s0.3 0.0933 b24/s0.3 0.0044 0.0000 dispersive-80ep n/a (no seed std)
Dispersive 80ep vs SiT 400K cknna cknna_dinov2 block8@s=0.1 0.0382 b8/s0.1 0.0118 b8/s0.1 0.0263 0.0000 dispersive-80ep n/a (no seed std)
Dispersive 80ep vs SiT 400K cknna cknna_dinov2 block8@s=0.3 0.0380 b8/s0.3 0.0107 b8/s0.3 0.0273 0.0000 dispersive-80ep n/a (no seed std)
Dispersive 80ep vs SiT 400K cknna cknna_dinov2 block8@s=0.5 0.0307 b8/s0.5 0.0099 b8/s0.5 0.0208 0.0000 dispersive-80ep n/a (no seed std)
Dispersive 80ep vs SiT 400K cknna cknna_dinov2 block8@s=0.7 0.0171 b8/s0.7 0.0076 b8/s0.7 0.0095 0.0000 dispersive-80ep n/a (no seed std)
Dispersive 80ep vs SiT 400K rank pr_token best(layer,s) 114.5039 b28/s0.1 145.0956 b28/s0.1 -30.5918 0.0000 n/a (no seed std)
Dispersive 80ep vs SiT 400K rank pr_token block8@s=0.1 40.2808 b8/s0.1 17.0594 b8/s0.1 23.2214 0.0000 n/a (no seed std)
Dispersive 80ep vs SiT 400K rank pr_token block8@s=0.3 39.7515 b8/s0.3 19.3032 b8/s0.3 20.4483 0.0000 n/a (no seed std)
Dispersive 80ep vs SiT 400K rank pr_token block8@s=0.5 40.1218 b8/s0.5 22.3664 b8/s0.5 17.7554 0.0000 n/a (no seed std)
Dispersive 80ep vs SiT 400K rank pr_token block8@s=0.7 40.8686 b8/s0.7 23.4953 b8/s0.7 17.3733 0.0000 n/a (no seed std)