Create quad_vae_liminal_results.txt

quad_vae_liminal_results.txt

@@ -0,0 +1,164 @@
+======================================================================
+Multi-VAE Geometric Comparison Pipeline
+======================================================================
+Downloading liminal.zip from AbstractPhil/grid-geometric-classifier-sliding-proto...
+Found 957 images
+
+======================================================================
+Step 1: Classifier
+======================================================================
+Loaded 638,387 params
+
+======================================================================
+Processing: SD 1.5
+  Model: stable-diffusion-v1-5/stable-diffusion-v1-5
+======================================================================
+  Latent: (4, 64, 64)  mean=0.014 std=3.719 [-33.188, 40.219]
+  Groups: [[0], [2], [3], [1]]
+  Scales: [(4, 64, 64), (4, 32, 32), (4, 16, 16)]
+SD 1.5 extract: 100%
+ 30/30 [00:07<00:00,  5.30×32/s]
+  SD 1.5: 105,895 annotations, conf=0.880
+    saddle                   59,460 ( 56.1%)
+    pentachoron              37,281 ( 35.2%)
+    triangular_prism          3,009 (  2.8%)
+    line_diag                 1,913 (  1.8%)
+    cone                      1,689 (  1.6%)
+
+======================================================================
+Processing: SDXL
+  Model: madebyollin/sdxl-vae-fp16-fix
+======================================================================
+config.json: 100%
+ 631/631 [00:00<00:00, 142kB/s]
+diffusion_pytorch_model.safetensors: 100%
+ 335M/335M [00:01<00:00, 329MB/s]
+Encoding: 100%
+ 240/240 [01:12<00:00,  1.65batch/s]
+  Latent: (4, 128, 128)  mean=0.605 std=5.059 [-18.938, 21.297]
+  Groups: [[0], [3], [2], [1]]
+  Scales: [(4, 64, 64), (4, 32, 32), (4, 16, 16)]
+SDXL extract: 100%
+ 30/30 [00:17<00:00,  2.13×32/s]
+  SDXL: 483,789 annotations, conf=0.872
+    saddle                  263,065 ( 54.4%)
+    pentachoron             137,450 ( 28.4%)
+    triangular_prism         26,717 (  5.5%)
+    line_diag                21,664 (  4.5%)
+    plane                    14,704 (  3.0%)
+
+======================================================================
+Processing: Flux.1
+  Model: black-forest-labs/FLUX.1-dev
+======================================================================
+  Latent: (16, 128, 128)  mean=-0.317 std=1.997 [-10.000, 7.500]
+  Groups: [[0, 6], [3, 2], [4, 14], [8, 5], [11, 13], [10, 7], [9, 15], [12, 1]]
+  Scales: [(16, 64, 64), (8, 32, 32), (8, 16, 16)]
+Flux.1 extract: 100%
+ 30/30 [00:44<00:00,  1.29s/×32]
+  Flux.1: 1,456,910 annotations, conf=0.891
+    plane                   527,329 ( 36.2%)
+    pentachoron             387,406 ( 26.6%)
+    square_xy               205,738 ( 14.1%)
+    saddle                  190,141 ( 13.1%)
+    triangular_prism         90,982 (  6.2%)
+
+======================================================================
+Processing: Flux.2
+  Model: black-forest-labs/FLUX.2-dev
+======================================================================
+Some weights of the model checkpoint at black-forest-labs/FLUX.2-dev were not used when initializing AutoencoderKL: 
+ ['bn.num_batches_tracked, bn.running_mean, bn.running_var']
+  Latent: (32, 128, 128)  mean=0.023 std=1.719 [-9.312, 9.812]
+  Groups: [[0, 20, 4, 26], [29, 15, 19, 2], [27, 3, 8, 9], [13, 14, 30, 24], [6, 22, 11, 18], [28, 12, 7, 1], [23, 17, 5, 10], [16, 31, 25, 21]]
+  Scales: [(16, 64, 64), (8, 32, 32), (8, 16, 16)]
+Flux.2 extract: 100%
+ 30/30 [00:58<00:00,  2.06s/×32]
+  Flux.2: 1,932,126 annotations, conf=0.871
+    saddle                1,295,564 ( 67.1%)
+    pentachoron             453,701 ( 23.5%)
+    tetrahedron              71,594 (  3.7%)
+    line_diag                49,691 (  2.6%)
+    triangular_prism         38,569 (  2.0%)
+
+======================================================================
+COMPARATIVE ANALYSIS: VAE Geometric Structures
+======================================================================
+
+──────────────────────────────────────────────────────────────────────
+  VAE                  Latent          Ann  Ann/img   Conf Classes
+──────────────────────────────────────────────────────────────────────
+  SD 1.5              4×64×64      105,895      111  0.880   18/38
+  SDXL              4×128×128      483,789      506  0.872   16/38
+  Flux.1           16×128×128    1,456,910     1522  0.891   24/38
+  Flux.2           32×128×128    1,932,126     2019  0.871   18/38
+
+──────────────────────────────────────────────────────────────────────
+TOP-5 CLASSES PER VAE
+──────────────────────��───────────────────────────────────────────────
+  SD 1.5        saddle:56%  pentachoron:35%  triangular_prism:3%  line_diag:2%  cone:2%
+  SDXL          saddle:54%  pentachoron:28%  triangular_prism:6%  line_diag:4%  plane:3%
+  Flux.1        plane:36%  pentachoron:27%  square_xy:14%  saddle:13%  triangular_prism:6%
+  Flux.2        saddle:67%  pentachoron:23%  tetrahedron:4%  line_diag:3%  triangular_prism:2%
+
+──────────────────────────────────────────────────────────────────────
+CLASS PRESENCE ACROSS VAEs (>0.5% of annotations)
+──────────────────────────────────────────────────────────────────────
+  Class                    SD 1.5       SDXL     Flux.1     Flux.2
+  saddle                    56.1%      54.4%      13.1%      67.1%
+  pentachoron               35.2%      28.4%      26.6%      23.5%
+  plane                     trace       3.0%      36.2%       0.8%
+  square_xy                 trace       1.5%      14.1%      trace
+  triangular_prism           2.8%       5.5%       6.2%       2.0%
+  tetrahedron                1.1%       1.7%       1.6%       3.7%
+  line_diag                  1.8%       4.5%      trace       2.6%
+  tube                       0.6%      trace       1.2%      trace
+  cone                       1.6%       0.7%      trace      trace
+
+──────────────────────────────────────────────────────────────────────
+GEOMETRIC FINGERPRINT SIMILARITY (cosine between class distributions)
+──────────────────────────────────────────────────────────────────────
+                   SD 1.5       SDXL     Flux.1     Flux.2
+  SD 1.5            1.000      0.994      0.521      0.974
+  SDXL              0.994      1.000      0.536      0.986
+  Flux.1            0.521      0.536      1.000      0.442
+  Flux.2            0.974      0.986      0.442      1.000
+
+──────────────────────────────────────────────────────────────────────
+DIMENSIONAL DISTRIBUTION
+──────────────────────────────────────────────────────────────────────
+  VAE                0D       1D       2D       3D   Curved
+  SD 1.5           0.0%     0.8%     0.5%    98.7%    42.3%
+  SDXL             0.0%     1.8%     2.5%    95.7%    42.0%
+  Flux.1           0.0%     0.8%    45.7%    53.5%    44.7%
+  Flux.2           0.0%     0.8%     1.0%    98.3%    63.9%
+
+──────────────────────────────────────────────────────────────────────
+CHANNEL GROUPS
+──────────────────────────────────────────────────────────────────────
+  SD 1.5       (4ch): [[0], [2], [3], [1]]
+  SDXL         (4ch): [[0], [3], [2], [1]]
+  Flux.1       (16ch): [[0, 6], [3, 2], [4, 14], [8, 5], [11, 13], [10, 7], [9, 15], [12, 1]]
+  Flux.2       (32ch): [[0, 20, 4, 26], [29, 15, 19, 2], [27, 3, 8, 9], [13, 14, 30, 24], [6, 22, 11, 18], [28, 12, 7, 1], [23, 17, 5, 10], [16, 31, 25, 21]]
+
+──────────────────────────────────────────────────────────────────────
+PER-IMAGE CROSS-VAE CONSISTENCY
+──────────────────────────────────────────────────────────────────────
+  Do images that are geometrically distinct in one VAE stay distinct in another?
+  SD 1.5       ↔ SDXL        : mean=0.940 std=0.084 [0.550, 1.000]
+    → Same geometric structure
+  SD 1.5       ↔ Flux.1      : mean=0.504 std=0.241 [0.007, 0.977]
+    → Different structures
+  SD 1.5       ↔ Flux.2      : mean=0.943 std=0.081 [0.583, 1.000]
+    → Same geometric structure
+  SDXL         ↔ Flux.1      : mean=0.527 std=0.241 [0.012, 0.979]
+    → Different structures
+  SDXL         ↔ Flux.2      : mean=0.956 std=0.076 [0.489, 1.000]
+    → Same geometric structure
+  Flux.1       ↔ Flux.2      : mean=0.461 std=0.251 [0.012, 0.973]
+    → Different structures
+
+Saved to /content/multi_vae_comparison.json
+======================================================================
+✓ Multi-VAE comparison complete!
+======================================================================