Create quad_vae_mega_liminal_results.txt

quad_vae_mega_liminal_results.txt

@@ -0,0 +1,168 @@
+======================================================================
+Multi-VAE Geometric Comparison Pipeline
+======================================================================
+Downloading mega_liminal_captioned.zip from AbstractPhil/grid-geometric-classifier-sliding-proto...
+mega_liminal_captioned.zip: 100%
+ 3.06G/3.06G [00:09<00:00, 224MB/s]
+Found 2074 images in /content/liminal
+
+======================================================================
+Step 1: Classifier
+======================================================================
+Loaded 638,387 params
+
+======================================================================
+Processing: SD 1.5
+  Model: stable-diffusion-v1-5/stable-diffusion-v1-5
+======================================================================
+Encoding: 100%
+ 418/418 [01:01<00:00,  2.51batch/s]
+  Latent: (4, 64, 64)  mean=-0.755 std=5.088 [-27.734, 34.625]
+  Groups: [[0], [2], [3], [1]]
+  Scales: [(4, 64, 64), (4, 32, 32), (4, 16, 16)]
+SD 1.5 extract: 100%
+ 65/65 [00:12<00:00,  5.62×32/s]
+  SD 1.5: 229,618 annotations, conf=0.880
+    saddle                  129,783 ( 56.5%)
+    pentachoron              80,065 ( 34.9%)
+    triangular_prism          6,523 (  2.8%)
+    line_diag                 4,241 (  1.8%)
+    cone                      3,630 (  1.6%)
+
+======================================================================
+Processing: SDXL
+  Model: madebyollin/sdxl-vae-fp16-fix
+======================================================================
+Encoding: 100%
+ 519/519 [02:28<00:00,  2.08batch/s]
+  Latent: (4, 128, 128)  mean=-0.372 std=6.734 [-23.984, 22.812]
+  Groups: [[0], [2], [3], [1]]
+  Scales: [(4, 64, 64), (4, 32, 32), (4, 16, 16)]
+SDXL extract: 100%
+ 65/65 [00:29<00:00,  2.27×32/s]
+  SDXL: 1,092,657 annotations, conf=0.874
+    saddle                  574,252 ( 52.6%)
+    pentachoron             332,100 ( 30.4%)
+    triangular_prism         66,206 (  6.1%)
+    line_diag                44,622 (  4.1%)
+    plane                    31,678 (  2.9%)
+
+======================================================================
+Processing: Flux.1
+  Model: black-forest-labs/FLUX.1-dev
+======================================================================
+Encoding: 100%
+ 519/519 [02:30<00:00,  2.01batch/s]
+  Latent: (16, 128, 128)  mean=-0.351 std=3.134 [-12.562, 10.500]
+  Groups: [[0, 6], [5, 2], [4, 9], [3, 7], [8, 10], [11, 13], [15, 1], [14, 12]]
+  Scales: [(16, 64, 64), (8, 32, 32), (8, 16, 16)]
+Flux.1 extract: 100%
+ 65/65 [01:09<00:00,  1.01×32/s]
+  Flux.1: 2,943,743 annotations, conf=0.878
+    pentachoron             906,158 ( 30.8%)
+    plane                   841,923 ( 28.6%)
+    saddle                  433,827 ( 14.7%)
+    square_xy               432,908 ( 14.7%)
+    triangular_prism        207,696 (  7.1%)
+
+======================================================================
+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.running_var, bn.num_batches_tracked, bn.running_mean']
+Encoding: 100%
+ 519/519 [02:32<00:00,  2.07batch/s]
+  Latent: (32, 128, 128)  mean=0.089 std=1.654 [-9.312, 9.312]
+  Groups: [[0, 20, 9, 18], [11, 2, 8, 19], [13, 30, 14, 12], [6, 7, 3, 5], [4, 17, 26, 24], [10, 25, 21, 22], [29, 15, 23, 1], [16, 31, 28, 27]]
+  Scales: [(16, 64, 64), (8, 32, 32), (8, 16, 16)]
+Flux.2 extract: 100%
+ 65/65 [01:37<00:00,  1.42s/×32]
+  Flux.2: 4,365,328 annotations, conf=0.875
+    saddle                3,040,696 ( 69.7%)
+    pentachoron             932,633 ( 21.4%)
+    tetrahedron             162,493 (  3.7%)
+    line_diag               125,249 (  2.9%)
+    triangular_prism         56,747 (  1.3%)
+
+======================================================================
+COMPARATIVE ANALYSIS: VAE Geometric Structures
+======================================================================
+
+──────────────────────────────────────────────────────────────────────
+  VAE                  Latent          Ann  Ann/img   Conf Classes
+──────────────────────────────────────────────────────────────────────
+  SD 1.5              4×64×64      229,618      111  0.880   18/38
+  SDXL              4×128×128    1,092,657      527  0.874   19/38
+  Flux.1           16×128×128    2,943,743     1419  0.878   24/38
+  Flux.2           32×128×128    4,365,328     2105  0.875   17/38
+
+─────────────────────────────────────────────��────────────────────────
+TOP-5 CLASSES PER VAE
+──────────────────────────────────────────────────────────────────────
+  SD 1.5        saddle:57%  pentachoron:35%  triangular_prism:3%  line_diag:2%  cone:2%
+  SDXL          saddle:53%  pentachoron:30%  triangular_prism:6%  line_diag:4%  plane:3%
+  Flux.1        pentachoron:31%  plane:29%  saddle:15%  square_xy:15%  triangular_prism:7%
+  Flux.2        saddle:70%  pentachoron:21%  tetrahedron:4%  line_diag:3%  triangular_prism:1%
+
+──────────────────────────────────────────────────────────────────────
+CLASS PRESENCE ACROSS VAEs (>0.5% of annotations)
+──────────────────────────────────────────────────────────────────────
+  Class                    SD 1.5       SDXL     Flux.1     Flux.2
+  saddle                    56.5%      52.6%      14.7%      69.7%
+  pentachoron               34.9%      30.4%      30.8%      21.4%
+  plane                     trace       2.9%      28.6%       0.8%
+  square_xy                 trace       1.6%      14.7%      trace
+  triangular_prism           2.8%       6.1%       7.1%       1.3%
+  tetrahedron                1.1%       1.4%       2.5%       3.7%
+  line_diag                  1.8%       4.1%      trace       2.9%
+  tube                       0.7%      trace       0.8%      trace
+  cone                       1.6%       0.6%      trace      trace
+
+──────────────────────────────────────────────────────────────────────
+GEOMETRIC FINGERPRINT SIMILARITY (cosine between class distributions)
+──────────────────────────────────────────────────────────────────────
+                   SD 1.5       SDXL     Flux.1     Flux.2
+  SD 1.5            1.000      0.996      0.615      0.966
+  SDXL              0.996      1.000      0.641      0.970
+  Flux.1            0.615      0.641      1.000      0.499
+  Flux.2            0.966      0.970      0.499      1.000
+
+──────────────────────────────────────────────────────────────────────
+DIMENSIONAL DISTRIBUTION
+──────────────────────────────────────────────────────────────────────
+  VAE                0D       1D       2D       3D   Curved
+  SD 1.5           0.0%     0.8%     0.5%    98.7%    42.8%
+  SDXL             0.0%     1.7%     2.3%    96.0%    42.9%
+  Flux.1           0.0%     0.6%    39.0%    60.4%    44.0%
+  Flux.2           0.0%     0.9%     0.9%    98.2%    65.3%
+
+──────────────────────────────────────────────────────────────────────
+CHANNEL GROUPS
+──────────────────────────────────────────────────────────────────────
+  SD 1.5       (4ch): [[0], [2], [3], [1]]
+  SDXL         (4ch): [[0], [2], [3], [1]]
+  Flux.1       (16ch): [[0, 6], [5, 2], [4, 9], [3, 7], [8, 10], [11, 13], [15, 1], [14, 12]]
+  Flux.2       (32ch): [[0, 20, 9, 18], [11, 2, 8, 19], [13, 30, 14, 12], [6, 7, 3, 5], [4, 17, 26, 24], [10, 25, 21, 22], [29, 15, 23, 1], [16, 31, 28, 27]]
+
+──────────────────────────────────────────────────────────────────────
+PER-IMAGE CROSS-VAE CONSISTENCY
+──────────────────────────────────────────────────────────────────────
+  Do images that are geometrically distinct in one VAE stay distinct in another?
+  SD 1.5       ↔ SDXL        : mean=0.916 std=0.102 [0.497, 1.000]
+    → Same geometric structure
+  SD 1.5       ↔ Flux.1      : mean=0.455 std=0.248 [0.019, 0.989]
+    → Different structures
+  SD 1.5       ↔ Flux.2      : mean=0.899 std=0.124 [0.345, 1.000]
+    → Similar structure
+  SDXL         ↔ Flux.1      : mean=0.517 std=0.232 [0.026, 0.987]
+    → Different structures
+  SDXL         ↔ Flux.2      : mean=0.905 std=0.128 [0.391, 0.999]
+    → Same geometric structure
+  Flux.1       ↔ Flux.2      : mean=0.378 std=0.245 [0.014, 0.986]
+    → Very different geometric encoding
+
+Saved to /content/multi_vae_comparison_mega.json
+======================================================================
+✓ Multi-VAE comparison complete!
+======================================================================