wiltodelta commited on
Commit
bcfc2f2
·
verified ·
1 Parent(s): 32a52ed

Source-classify freeze (GitHub 9df5c2d)

Browse files
README.md CHANGED
@@ -8,81 +8,84 @@ tags:
8
  - open-set-classification
9
  ---
10
 
11
- # RAIW multi-provider source-pipeline classifier (`multi-provider-source-v3`)
12
 
13
- This repository contains the pickle-free (`allow_pickle=False`) derivative weights for the abstaining `remove-ai-watermarks` pixel source-pipeline classifier across **10 AI provider families** (`openai`, `google`, `bytedance`, `flux`, `stable_diffusion`, `meta`, `midjourney`, `ideogram`, `firefly`, `xai`) plus `unknown` (`classify_source` / `remove-ai-watermarks classify-source`).
14
 
15
  This model is **not a SynthID detector** and does not decode or verify any cryptographic or invisible watermark payload (`watermark_truth="unknown"`). Its labels describe complete generation and export pixel pipelines, so renderer, geometry, encoding, post-processing, and watermark state may be confounded. `unknown` does not mean that SynthID is absent, and a provider label does not prove that SynthID is present.
16
 
17
  ## Model architecture
18
 
19
- - **Input (`956-d`):**
20
- - `124-d` spatial opponent-residual and gradient forensic features (`extract_124d`).
21
- - `768-d` phase-free `16x16` log-power and radial spectral features (`source_spectral`).
22
- - `64-d` directional / block-boundary (`8x8` and `16x16`) / chroma-aliasing / azimuthal-anisotropy features (`augment_892d_to_956d`).
23
- - **Stage 1 (`54` primary specialist units + `1,590` auxiliary hard-margin subfamily, codec-recompression, cross-AI disambiguator, and domain-veto hinge descriptors across `10` canonical providers):**
24
- - `10` global provider specialists + `44` subfamily and codec/export-regime specialists (`37` active), each combining a block-diagonal MLP Soup (including 3-seed joint multi-task `54`-unit shared-trunk MLP blocks), a vectorized fixed-depth `HistGradientBoosting` tree ensemble, and a distilled `ExtraTrees -> MLP` student with calibration-reoptimized distilled blend weights (`w_e <= 0.25`), pooled via drift-bounded zero-inflation subfamily blending across 16 real-image varieties, refined by a Wave-24 Stage-1.5 2-knot zero-inflated hinge meta-router (`router_meta_tau`, `router_meta_w`) incorporating expanded-corpus tail-purified 3-component hybrid (`MLP Soup + HGB + Distilled ExtraTrees-MLP`), pre-ReLU extreme-tail zero-FP composite blends (`_tailmax`), pure-channel (`onlyhgb`, `onlymlp`, `onlyet`), extreme-value-safe steep zero-FP tail hinges (`SteepHinge`), provenance-cleaned cross-AI disambiguator (`Disambig` / `ParetoDis`), and real-domain veto specialists, and routed by a unified damped competitive margin (`gamma = 0.60`).
25
- - **Stage 2 (`9` stacked version heads, `2600-d`):**
26
- - Feature-subsampled vectorized `HistGradientBoosting` tree ensembles (including multi-seed/multi-depth forest soups on `openai`, `google`, `flux`, `midjourney`, `ideogram`, `bytedance`, and `stable_diffusion`) taking the `956-d` features (and Stage-1 unit log-odds scores) to predict the canonical model/export subfamily (`version_subfamily`).
27
- - **Artifact:** `multi-provider-source-v3.npz` (`allow_pickle=False`)
28
- - **SHA-256:** `f65dac7cd7a6d8025e0b849bba9117635d6280628c226f47b599332898580f3c`
 
 
 
29
 
30
  The artifact contains only learned parameters, normalization constants, tree arrays, labels, and calibrated thresholds. It does not contain training images, paths, image hashes, embeddings, or a training catalog.
31
 
32
- ## Evaluation (`204,050` clean audited samples; `10,586` locked test)
33
 
34
- The operating point was selected on a disjoint `8,365`-image calibration split across 16 fine-grained real-image varieties and group-locked (`parent_sha256`) native, light-JPEG, and heavy-JPEG/resize views. On the `10,586`-image locked test split (`4,539` multi-variety `unknown` negatives + `6,047` AI provider images including preprocessed social/JPEG derivatives):
35
 
36
- - **Unified 10-provider router:** `99.52%` `unknown` specificity (`0.48%` test `unknown` FPR; `0.73%` photographic FPR and `0.25%` non-photographic structured FPR across all 15 non-photo varieties), `69.97%` aggregate target recall across native (`76.72%`) and recompressed (`35.74%`) test views (`98.42%` provider accuracy among classified AI; `97.92%` overall precision when firing), and `0.9577` mean standalone ROC-AUC / `0.9682` binary PR-AUC on the monotonic `confidence` score (`_sigmoid_scalar(excess[best_cls])`).
37
 
38
  | Provider | Test count | Standalone ROC AUC | Router recall | Router precision | Stage-2 version accuracy |
39
  |---|---:|---:|---:|---:|---:|
40
- | `openai` | 1,478 | 0.9740 | 77.3% | 97.9% | 86.5% |
41
- | `google` | 1,382 | 0.9759 | 76.2% | 97.5% | 93.7% |
42
- | `bytedance` | 389 | 0.9582 | 68.9% | 99.6% | 92.0% |
43
- | `flux` | 1,357 | 0.9505 | 59.4% | 98.3% | 87.1% |
44
- | `stable_diffusion` | 955 | 0.9742 | 71.5% | 98.0% | 90.8% |
45
- | `meta` | 94 | 0.9806 | 81.9% | 100.0% | — (`meta:imagine`) |
46
- | `midjourney` | 123 | 0.9594 | 57.7% | 94.7% | 91.1% |
47
- | `ideogram` | 118 | 0.9361 | 47.5% | 98.2% | 65.3% |
48
- | `firefly` | 103 | 0.9243 | 56.3% | 93.5% | 100.0% |
49
- | `xai` | 48 | 0.9438 | 35.4% | 100.0% | 72.9% |
50
-
51
- ### Held-out real-image specificity by variety (`4,539` test negatives across 16 real classes + foreign AI)
52
 
53
  | Real-image variety | Test count | False positives | False-positive rate | Specificity |
54
  |---|---:|---:|---:|---:|
55
- | `photo` | 2,183 | 16 | 0.73% | 99.27% |
56
- | `photo_retouched` | 339 | 2 | 0.59% | 99.41% |
57
  | `classical_inpaint` | 17 | 0 | 0.00% | 100.00% |
58
- | `digital_art` | 441 | 1 | 0.23% | 99.77% |
59
- | `game_art` | 174 | 0 | 0.00% | 100.00% |
60
- | `vector` | 135 | 0 | 0.00% | 100.00% |
61
  | `interface` | 180 | 0 | 0.00% | 100.00% |
62
- | `diagram` | 98 | 1 | 1.02% | 98.98% |
63
  | `chart_data_viz` | 159 | 0 | 0.00% | 100.00% |
64
  | `technical_architecture` | 84 | 0 | 0.00% | 100.00% |
65
  | `procedural_graphic` | 13 | 0 | 0.00% | 100.00% |
66
  | `document` | 165 | 0 | 0.00% | 100.00% |
67
  | `gov_scientific_doc` | 32 | 0 | 0.00% | 100.00% |
68
  | `receipt_invoice` | 181 | 0 | 0.00% | 100.00% |
69
- | `print_publishing` | 229 | 1 | 0.44% | 99.56% |
70
- | `stress_recompressed` | 103 | 1 | 0.97% | 99.03% |
71
  | `foreign_ai` | 6 | 0 | 0.00% | 100.00% |
72
 
73
  ### External blind multi-variety real-image verification (`selective-v8`, `2,267` untouched images)
74
 
75
- Evaluated on the hash-disjoint external `selective-v8` blind real-image cohorts (`2,267` images never seen in train, calibration, or test): **2 / 2,267 false positives (`0.09%` overall FPR, `99.91%` specificity)**.
76
 
77
  | Blind `selective-v8` cell | Evaluated | False positives | False-positive rate | Specificity |
78
  |---|---:|---:|---:|---:|
79
  | `photo` | 450 | 1 | 0.22% | 99.78% |
80
  | `vector` | 332 | 0 | 0.00% | 100.00% |
81
- | `game_art` | 230 | 1 | 0.43% | 99.57% |
82
  | `interface` | 342 | 0 | 0.00% | 100.00% |
83
  | `diagram` | 305 | 0 | 0.00% | 100.00% |
84
  | `document` | 308 | 0 | 0.00% | 100.00% |
85
- | `stress_recompressed` | 300 | 0 | 0.00% | 100.00% |
86
 
87
  ## Intended use
88
 
 
8
  - open-set-classification
9
  ---
10
 
11
+ # RAIW multi-provider source-pipeline classifier (`multi-provider-source-v4`)
12
 
13
+ This repository contains the compact (`13.35 MB`), pickle-free (`allow_pickle=False`) derivative weights for the abstaining `remove-ai-watermarks` pixel source-pipeline classifier across **10 AI provider families** (`openai`, `google`, `bytedance`, `flux`, `stable_diffusion`, `meta`, `midjourney`, `ideogram`, `firefly`, `xai`) plus `unknown` (`classify_source` / `remove-ai-watermarks classify-source`).
14
 
15
  This model is **not a SynthID detector** and does not decode or verify any cryptographic or invisible watermark payload (`watermark_truth="unknown"`). Its labels describe complete generation and export pixel pipelines, so renderer, geometry, encoding, post-processing, and watermark state may be confounded. `unknown` does not mean that SynthID is absent, and a provider label does not prove that SynthID is present.
16
 
17
  ## Model architecture
18
 
19
+ - **Input (`512-d`, `extract_512d_v4_features`):**
20
+ - `144-d` multi-scale spatial SRM residuals, directional kurtosis, and cross-channel Y/Co/Cg co-occurrence statistics.
21
+ - `128-d` phase-invariant modulo-8 / modulo-16 sublattice energy, VAE upsampler checkerboards, `8x8` block-DCT quantization combs, and 4:2:0 chroma block-boundary ratios.
22
+ - `128-d` phase-free multi-patch radial and azimuthal log-power spectra, high-frequency rolloff slopes, and resampling second-derivative spectral peaks.
23
+ - `80-d` camera-sensor physics (Bayer CFA demosaicing periodicity, local noise-level heteroscedasticity, highlight clipping, edge ringing, and document/text binarization cues).
24
+ - `32-d` explicit codec, aspect-ratio, scale, and social-recompression regime vector (`x[480:512]`).
25
+ - **Stage 1 (`132` unit channels + `68` 2D multiplicative interaction gates across `10` canonical providers):**
26
+ - Two-tower ensemble executed in pure NumPy: `5` multi-seed Regime-FiLM `LayerNorm` + `SiLU` residual MLP trunks (`11` provider/unknown logits + `45` subfamily logits) paired with `44` packed `HistGradientBoosting` specialist forests (global provider, subfamily, codec-regime, hard-domain veto, and pairwise cross-AI disambiguator heads).
27
+ - Smooth softplus hinge activations combined with `68` 2D multiplicative interaction gates (`provider x hard-domain veto`, `provider x pairwise disambiguator`, `MLP x HGB` agreement), standardized against a learned regime-conditioned null baseline `(z_raw - mu(r)) / sigma(r)` and routed via damped competitive margins ($\gamma = 0.18$).
28
+ - **Stage 2 (`9` stacked version heads, `644-d`):**
29
+ - Vectorized `HistGradientBoosting` tree ensembles taking the `512-d` descriptor concatenated with the `132` Stage-1 unit scores to predict the canonical model/export subfamily (`version_subfamily`).
30
+ - **Artifact:** `multi-provider-source-v4.npz` (`allow_pickle=False`, `13.35 MB`)
31
+ - **SHA-256:** `bc62ee22331c0137e76c53f5026e14d08cae12a3483fa69f3000316806c146a3`
32
 
33
  The artifact contains only learned parameters, normalization constants, tree arrays, labels, and calibrated thresholds. It does not contain training images, paths, image hashes, embeddings, or a training catalog.
34
 
35
+ ## Evaluation (`127,001` deduplicated feature views; `10,585` locked test)
36
 
37
+ The operating point was selected strictly on a disjoint calibration partition (`8,363` primary calibration views + `3,011` social-recompressed real negatives across 16 fine-grained real-image varieties and group-locked `parent_sha256` families). On the `10,585`-image locked test split (`4,538` multi-variety `unknown` negatives + `6,047` AI provider images across native, resized, and social-recompressed views):
38
 
39
+ - **Unified 10-provider router:** `99.76%` `unknown` specificity (`0.24%` test `unknown` FPR; `0.14%` photographic FPR and `0.34%` non-photographic structured FPR), `81.35%` exact provider recall (`85.23%` any-AI recall; `88.08%` native, `84.58%` resized, and `76.81%` recompressed exact recall), `99.79%` any-AI precision when firing (`95.44%` exact-provider accuracy among classified AI; `96.32%` overall Stage-2 version accuracy given exact provider), and `0.9974` binary ROC-AUC / `0.9982` binary PR-AUC (`0.9976` mean standalone provider ROC-AUC).
40
 
41
  | Provider | Test count | Standalone ROC AUC | Router recall | Router precision | Stage-2 version accuracy |
42
  |---|---:|---:|---:|---:|---:|
43
+ | `openai` | 1,478 | 0.9997 | 92.5% | 96.4% | 95.8% |
44
+ | `google` | 1,382 | 0.9993 | 85.5% | 96.3% | 97.3% |
45
+ | `bytedance` | 389 | 0.9853 | 69.9% | 98.9% | 97.8% |
46
+ | `flux` | 1,357 | 0.9992 | 80.6% | 94.2% | 96.2% |
47
+ | `stable_diffusion` | 955 | 0.9990 | 70.5% | 95.3% | 97.8% |
48
+ | `meta` | 94 | 0.9987 | 89.4% | 93.3% | - (`meta:imagine`) |
49
+ | `midjourney` | 123 | 1.0000 | 63.4% | 92.9% | 100.0% |
50
+ | `ideogram` | 118 | 0.9997 | 75.4% | 73.0% | 74.2% |
51
+ | `firefly` | 103 | 0.9959 | 58.3% | 96.8% | 100.0% |
52
+ | `xai` | 48 | 0.9996 | 43.8% | 100.0% | 71.4% |
53
+
54
+ ### Held-out real-image specificity by variety (`4,538` test negatives across 16 real classes + foreign AI)
55
 
56
  | Real-image variety | Test count | False positives | False-positive rate | Specificity |
57
  |---|---:|---:|---:|---:|
58
+ | `photo` | 2,183 | 3 | 0.14% | 99.86% |
59
+ | `photo_retouched` | 339 | 1 | 0.29% | 99.71% |
60
  | `classical_inpaint` | 17 | 0 | 0.00% | 100.00% |
61
+ | `digital_art` | 440 | 1 | 0.23% | 99.77% |
62
+ | `game_art` | 174 | 3 | 1.72% | 98.28% |
63
+ | `vector` | 135 | 1 | 0.74% | 99.26% |
64
  | `interface` | 180 | 0 | 0.00% | 100.00% |
65
+ | `diagram` | 98 | 0 | 0.00% | 100.00% |
66
  | `chart_data_viz` | 159 | 0 | 0.00% | 100.00% |
67
  | `technical_architecture` | 84 | 0 | 0.00% | 100.00% |
68
  | `procedural_graphic` | 13 | 0 | 0.00% | 100.00% |
69
  | `document` | 165 | 0 | 0.00% | 100.00% |
70
  | `gov_scientific_doc` | 32 | 0 | 0.00% | 100.00% |
71
  | `receipt_invoice` | 181 | 0 | 0.00% | 100.00% |
72
+ | `print_publishing` | 229 | 0 | 0.00% | 100.00% |
73
+ | `stress_recompressed` | 103 | 2 | 1.94% | 98.06% |
74
  | `foreign_ai` | 6 | 0 | 0.00% | 100.00% |
75
 
76
  ### External blind multi-variety real-image verification (`selective-v8`, `2,267` untouched images)
77
 
78
+ Evaluated on the hash-disjoint external `selective-v8` blind real-image cohorts (`2,267` images never seen in train, calibration, or test): **2 / 2,267 false positives (`0.09%` overall FPR, `99.91%` specificity)**, plus **0 / 60 false positives (`0.00%` FPR)** on external phone thermal receipts.
79
 
80
  | Blind `selective-v8` cell | Evaluated | False positives | False-positive rate | Specificity |
81
  |---|---:|---:|---:|---:|
82
  | `photo` | 450 | 1 | 0.22% | 99.78% |
83
  | `vector` | 332 | 0 | 0.00% | 100.00% |
84
+ | `game_art` | 230 | 0 | 0.00% | 100.00% |
85
  | `interface` | 342 | 0 | 0.00% | 100.00% |
86
  | `diagram` | 305 | 0 | 0.00% | 100.00% |
87
  | `document` | 308 | 0 | 0.00% | 100.00% |
88
+ | `stress_recompressed` | 300 | 1 | 0.33% | 99.67% |
89
 
90
  ## Intended use
91
 
metrics.json CHANGED
@@ -1,10 +1,10 @@
1
  {
2
  "claim": "abstaining 10-provider source-pipeline classification with version subfamily attribution; not SynthID detection",
3
- "schema": "multi-provider-source-v3",
4
- "artifact_file": "multi-provider-source-v3.npz",
5
- "artifact_sha256": "f65dac7cd7a6d8025e0b849bba9117635d6280628c226f47b599332898580f3c",
6
- "feature_dimension": 956,
7
- "stacked_stage2_dimension": 2600,
8
  "classes": [
9
  "openai",
10
  "google",
@@ -19,61 +19,63 @@
19
  "unknown"
20
  ],
21
  "corpus_counts": {
22
- "total": 204050,
23
- "train": 185099,
24
- "calibration": 8365,
25
- "test": 10586
 
26
  },
27
  "locked_internal_test": {
28
- "n": 10586,
29
- "binary_roc_auc": 0.9577,
30
- "binary_pr_auc": 0.9682,
31
- "unknown_specificity": 0.995153117426746,
32
- "unknown_fpr": 0.004846882573254021,
33
- "photo_fpr": 0.007329363261566651,
34
- "structured_fpr": 0.0025466893039049238,
35
- "aggregate_target_recall": 0.699685794608897,
 
36
  "recall": {
37
- "openai": 0.7726657645466847,
38
- "google": 0.7619392185238785,
39
- "bytedance": 0.6889460154241646,
40
- "flux": 0.5939572586588062,
41
- "stable_diffusion": 0.7151832460732984,
42
- "meta": 0.8191489361702128,
43
- "midjourney": 0.5772357723577236,
44
- "ideogram": 0.4745762711864407,
45
- "firefly": 0.5631067961165048,
46
- "xai": 0.3541666666666667
47
  },
48
  "precision": {
49
- "openai": 0.9785775492716366,
50
- "google": 0.975,
51
- "bytedance": 0.9962825278810409,
52
- "flux": 0.9829268292682927,
53
- "stable_diffusion": 0.9799139167862266,
54
- "meta": 1.0,
55
- "midjourney": 0.9466666666666667,
56
- "ideogram": 0.9824561403508771,
57
- "firefly": 0.9354838709677419,
58
  "xai": 1.0
59
  },
60
  "real_variety_specificity": {
61
- "photo": 0.9926706367384334,
62
- "photo_retouched": 0.9941002949852508,
63
  "classical_inpaint": 1.0,
64
- "digital_art": 0.9977324263038548,
65
- "game_art": 1.0,
66
- "vector": 1.0,
67
  "interface": 1.0,
68
- "diagram": 0.9897959183673469,
69
  "chart_data_viz": 1.0,
70
  "technical_architecture": 1.0,
71
  "procedural_graphic": 1.0,
72
  "document": 1.0,
73
  "gov_scientific_doc": 1.0,
74
  "receipt_invoice": 1.0,
75
- "print_publishing": 0.9956331877729258,
76
- "stress_recompressed": 0.9902912621359223,
77
  "foreign_ai": 1.0
78
  }
79
  },
@@ -97,9 +99,9 @@
97
  },
98
  "game_art": {
99
  "evaluated": 230,
100
- "false_positives": 1,
101
- "fpr": 0.004347826086956522,
102
- "specificity": 0.9956521739130435
103
  },
104
  "interface": {
105
  "evaluated": 342,
@@ -121,21 +123,21 @@
121
  },
122
  "stress_recompressed": {
123
  "evaluated": 300,
124
- "false_positives": 0,
125
- "fpr": 0.0,
126
- "specificity": 1.0
127
  }
128
  }
129
  },
130
  "stage2_version_accuracy": {
131
- "openai": 0.86468200270636,
132
- "google": 0.9370477568740955,
133
- "bytedance": 0.9203084832904884,
134
- "flux": 0.871039056742815,
135
- "stable_diffusion": 0.9078534031413612,
136
- "midjourney": 0.9105691056910569,
137
- "ideogram": 0.652542372881356,
138
  "firefly": 1.0,
139
- "xai": 0.7291666666666666
140
  }
141
  }
 
1
  {
2
  "claim": "abstaining 10-provider source-pipeline classification with version subfamily attribution; not SynthID detection",
3
+ "schema": "multi-provider-source-v4",
4
+ "artifact_file": "multi-provider-source-v4.npz",
5
+ "artifact_sha256": "bc62ee22331c0137e76c53f5026e14d08cae12a3483fa69f3000316806c146a3",
6
+ "feature_dimension": 512,
7
+ "stacked_stage2_dimension": 644,
8
  "classes": [
9
  "openai",
10
  "google",
 
19
  "unknown"
20
  ],
21
  "corpus_counts": {
22
+ "total": 127001,
23
+ "train": 105042,
24
+ "calibration": 8363,
25
+ "calibration_recomp_neg": 3011,
26
+ "test": 10585
27
  },
28
  "locked_internal_test": {
29
+ "n": 10585,
30
+ "binary_roc_auc": 0.9974117831066663,
31
+ "binary_pr_auc": 0.9981946495110887,
32
+ "unknown_specificity": 0.997576024680476,
33
+ "unknown_fpr": 0.0024239753195240193,
34
+ "photo_fpr": 0.001374255611543747,
35
+ "structured_fpr": 0.003397027600849257,
36
+ "aggregate_target_recall": 0.8134612204398876,
37
+ "any_ai_recall": 0.8523234661815776,
38
  "recall": {
39
+ "openai": 0.9248985115020297,
40
+ "google": 0.8545586107091172,
41
+ "bytedance": 0.699228791773779,
42
+ "flux": 0.8061901252763449,
43
+ "stable_diffusion": 0.7047120418848167,
44
+ "meta": 0.8936170212765957,
45
+ "midjourney": 0.6341463414634146,
46
+ "ideogram": 0.7542372881355932,
47
+ "firefly": 0.5825242718446602,
48
+ "xai": 0.4375
49
  },
50
  "precision": {
51
+ "openai": 0.964033850493653,
52
+ "google": 0.9632952691680261,
53
+ "bytedance": 0.9890909090909091,
54
+ "flux": 0.94229112833764,
55
+ "stable_diffusion": 0.953257790368272,
56
+ "meta": 0.9333333333333333,
57
+ "midjourney": 0.9285714285714286,
58
+ "ideogram": 0.7295081967213115,
59
+ "firefly": 0.967741935483871,
60
  "xai": 1.0
61
  },
62
  "real_variety_specificity": {
63
+ "photo": 0.9986257443884563,
64
+ "photo_retouched": 0.9970501474926253,
65
  "classical_inpaint": 1.0,
66
+ "digital_art": 0.9977272727272727,
67
+ "game_art": 0.9827586206896551,
68
+ "vector": 0.9925925925925926,
69
  "interface": 1.0,
70
+ "diagram": 1.0,
71
  "chart_data_viz": 1.0,
72
  "technical_architecture": 1.0,
73
  "procedural_graphic": 1.0,
74
  "document": 1.0,
75
  "gov_scientific_doc": 1.0,
76
  "receipt_invoice": 1.0,
77
+ "print_publishing": 1.0,
78
+ "stress_recompressed": 0.9805825242718447,
79
  "foreign_ai": 1.0
80
  }
81
  },
 
99
  },
100
  "game_art": {
101
  "evaluated": 230,
102
+ "false_positives": 0,
103
+ "fpr": 0.0,
104
+ "specificity": 1.0
105
  },
106
  "interface": {
107
  "evaluated": 342,
 
123
  },
124
  "stress_recompressed": {
125
  "evaluated": 300,
126
+ "false_positives": 1,
127
+ "fpr": 0.0033333333333333335,
128
+ "specificity": 0.9966666666666667
129
  }
130
  }
131
  },
132
  "stage2_version_accuracy": {
133
+ "openai": 0.958302852962692,
134
+ "google": 0.9729043183742591,
135
+ "bytedance": 0.9779411764705882,
136
+ "flux": 0.9616087751371115,
137
+ "stable_diffusion": 0.9777117384843982,
138
+ "midjourney": 1.0,
139
+ "ideogram": 0.7415730337078652,
140
  "firefly": 1.0,
141
+ "xai": 0.7142857142857143
142
  }
143
  }
multi-provider-source-v3.npz → multi-provider-source-v4.npz RENAMED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:f65dac7cd7a6d8025e0b849bba9117635d6280628c226f47b599332898580f3c
3
- size 2030769977
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:bc62ee22331c0137e76c53f5026e14d08cae12a3483fa69f3000316806c146a3
3
+ size 13996740