yomi-chen commited on
Commit
6cabe1b
·
verified ·
1 Parent(s): 94c6854

Clarify independent-size MLLM inputs and add16 real Gemini JoyAI checks

Browse files
README.md CHANGED
@@ -19,7 +19,7 @@ size_categories:
19
 
20
  EditableDesignBench **v4.0.0 / run3** 的六个模型结果:每模型 **文字 100 张 + 图片资产 100 张**,共 **1,200 张**。
21
 
22
- 全部图片与 MSE、PSNR、SSIM、LPIPS 已完成。**尚未运行 MLLM 评分。**
23
 
24
  [数据集](https://huggingface.co/datasets/yomi-chen/EditableDesignBench) · [评测代码](https://github.com/Ch3nSir/EditableDesignBench)
25
 
@@ -92,17 +92,36 @@ python evaluation_code/scripts/mllm_evaluate.py \
92
  --workers 10 --require-scoring
93
  ```
94
 
95
- 输入为完整原图、完整结果图和自然语言指令,不提供 mask 或目标图给评委。每例一次请求返回四维评分;重复命令会复用已成功的评分,只补失败/缺失项。本仓库发布时没有进行上述 API 调用。
96
 
97
- ### JoyAI 的尺寸区别
98
 
99
- JoyAI 使用官方默认输出档位(1024-base bucket,30 steps,CFG 4),直接输入原图,不指定输出宽高。输出可能与原图尺寸不同;完整变换记录在每条记录的 `official_geometry` 中。
100
 
101
- **JoyAI 也可直接使用上述 MLLM 命令**:将 `--predictions` 替换为 `text/joyai-image-edit-plus/predictions` 或 `assets/joyai-image-edit-plus/predictions`,资产轨同时使用 `--dataset data/EditableDesignBench-v4/assets`。新代码已去掉 MLLM 的同尺寸要求;输入为完整原图 + 完整官方输出 + 原始指令,不做缩放、裁剪、补边或 Alpha 合成。这个入口不读取历史生成报告。
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
102
 
103
- 不同尺寸不会直接导致评分失败;是否丢失内容、破坏布局仍由评委按原有 rubric 正常判断。像素指标的规则独立:已发布的 JoyAI 像素分数仍使用匹配官方几何变换的参考图与 mask,预测图没有缩放,所以其像素分数应单独解读。
 
 
 
104
 
105
- 如果已下载旧版结果,只需重新下载 `archives/evaluation_code.zip` 并解压覆盖 `evaluation_code/`,不必重新下载模型图片。本次兼容性修复未调用评委 API,也未改变已有图片或像素分数。
106
 
107
  ## 📊 指标与版本
108
 
 
19
 
20
  EditableDesignBench **v4.0.0 / run3** 的六个模型结果:每模型 **文字 100 张 + 图片资产 100 张**,共 **1,200 张**。
21
 
22
+ 全部图片与 MSE、PSNR、SSIM、LPIPS 已完成。**尚未进行六个模型的全量 MLLM 评分。** 小样本 API 检查与全量分数分开记录。
23
 
24
  [数据集](https://huggingface.co/datasets/yomi-chen/EditableDesignBench) · [评测代码](https://github.com/Ch3nSir/EditableDesignBench)
25
 
 
92
  --workers 10 --require-scoring
93
  ```
94
 
95
+ 输入为完整原图、完整结果图和自然语言指令,不提供 mask 或目标图给评委。每例一次请求返回四维评分;重复命令会复用已成功的评分,只补失败/缺失项。上述命令会实际调用使用者配置的评委 API;发布的模型图片不需要重新生成。
96
 
97
+ ### JoyAI:MLLM 与像素指标的尺寸规则不同
98
 
99
+ **MLLM 可直接接收不同尺寸的两张整图。** JoyAI 使用官方默认档位(1024-base bucket,30 steps,CFG 4);例如原图 1600 × 1200、输出 1152 × 832,评委接收两张各自完整的图与指令,不缩放、不裁剪、不补边、不进行 Alpha 合成:
100
 
101
+ ```bash
102
+ python evaluation_code/scripts/mllm_evaluate.py \
103
+ --dataset data/EditableDesignBench-v4 \
104
+ --predictions text/joyai-image-edit-plus/predictions \
105
+ --output judge/text/joyai-image-edit-plus.json \
106
+ --workers 10 --require-scoring
107
+ ```
108
+
109
+ 资产轨同时将数据目录改为 `data/EditableDesignBench-v4/assets`、图片目录改为 `assets/joyai-image-edit-plus/predictions`。不同尺寸不会直接导致拒收;内容丢失或布局变化仍按原有 rubric 正常评分。
110
+
111
+ **像素指标需要对应的评测坐标。** 已发布的 JoyAI 像素分数将原图和 mask 按官方比例缩放、中心裁剪后,与未缩放的结果图比较。变换记录在 `official_geometry` 中;被裁掉的边缘不参与像素计算。因此它与其他模型的原尺寸像素分数应分开解读。
112
+
113
+ 已下载旧包的用户只需重新下载 `archives/evaluation_code.zip`,解压覆盖 `evaluation_code/`;图片与既有像素指标均未改变。
114
+
115
+ ### ✅ Gemini API 小样本实测
116
+
117
+ 使用 Gemini-4-Argon 实测 JoyAI 的文字 8 条、资产 8 条,覆盖两轨全部编辑操作。两轨均为 **8/8 成功**,共 64 个维度分数,完整原图和官方输出保持不同尺寸直接输入。每例一次请求返回四维评分,网络或格式错误会按缓存机制重试。
118
 
119
+ | 轨道 | 成功 | IF | TA / OA | VC | LP |
120
+ |---|---:|---:|---:|---:|---:|
121
+ | 文字 | 8/8 | 2.250 | 3.000 | 2.875 | 2.000 |
122
+ | 资产 | 8/8 | 2.125 | 2.375 | 2.875 | 2.125 |
123
 
124
+ 含重试共 **19 次 API 调用**。这 16 条仅验证接口与输入流程,不能作为模型全量排名;六个模型的全量 MLLM 分数仍未提供。逐例分数与抽样任务见 [实测记录](validation/joyai_gemini16/README.md)。
125
 
126
  ## 📊 指标与版本
127
 
files_manifest.json CHANGED
@@ -12,8 +12,8 @@
12
  },
13
  {
14
  "path": "README.md",
15
- "bytes": 6269,
16
- "sha256": "8c718f82486227ada4835753956cf0da6b87c7c65b6404d4ea8d7fea251f3c38"
17
  },
18
  {
19
  "path": "archives/EditableDesignBench_v4_data.zip",
@@ -87,8 +87,28 @@
87
  },
88
  {
89
  "path": "release_info.json",
90
- "bytes": 2210,
91
- "sha256": "ec3ab87bc0cda990e18f66f4732420e823ebc81e10597b9496f17b08ed064c7f"
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
92
  }
93
  ],
94
  "images": 1200,
 
12
  },
13
  {
14
  "path": "README.md",
15
+ "bytes": 7236,
16
+ "sha256": "75184aa44c0f9a9d40d5cea1e22c58a40e63f6f34cf3dafc43779085936c5064"
17
  },
18
  {
19
  "path": "archives/EditableDesignBench_v4_data.zip",
 
87
  },
88
  {
89
  "path": "release_info.json",
90
+ "bytes": 2421,
91
+ "sha256": "4b6008ca3065ce1ac5f8d8a0bc3723ad7028f3ec0233d4badae53dc60500d1f9"
92
+ },
93
+ {
94
+ "path": "validation/joyai_gemini16/README.md",
95
+ "bytes": 515,
96
+ "sha256": "e1451dfbf4930faf7d896696d511d9a0782a069a5389bedb43301b9795113aa9"
97
+ },
98
+ {
99
+ "path": "validation/joyai_gemini16/case_scores.csv",
100
+ "bytes": 6303,
101
+ "sha256": "2d3aeeefdd71ad41aaab89166a81a8a7540e03996612aab6ee65e6ca6b0fb8ae"
102
+ },
103
+ {
104
+ "path": "validation/joyai_gemini16/selection.json",
105
+ "bytes": 4418,
106
+ "sha256": "0c92cb734bf29ad9100c7f073429692039a9c83aac5ae7c255cc586d473afca5"
107
+ },
108
+ {
109
+ "path": "validation/joyai_gemini16/summary.json",
110
+ "bytes": 16571,
111
+ "sha256": "aa9a075dee912d3924b4ae7159ff749e70e808b22382ded6ef71b4f181940cbd"
112
  }
113
  ],
114
  "images": 1200,
release_info.json CHANGED
@@ -40,5 +40,13 @@
40
  "reports_note": "recorded_reports preserve historical reports with workspace root paths replaced by logical labels; they are reference records, not resumable caches.",
41
  "evaluation_code_commit": "2dc67b321c948d1976dc7bf5c00910e7cd41060e",
42
  "mllm_image_size_policy": "External source and prediction may differ in size; both full RGB images are sent unchanged, no crop/resize/padding/compositing. Pixel protocols are unchanged.",
43
- "evaluation_code_note": "Frozen generation/runtime snapshot with the independently sized MLLM input compatibility fix; original source_commit/code_commit identifies generation."
 
 
 
 
 
 
 
 
44
  }
 
40
  "reports_note": "recorded_reports preserve historical reports with workspace root paths replaced by logical labels; they are reference records, not resumable caches.",
41
  "evaluation_code_commit": "2dc67b321c948d1976dc7bf5c00910e7cd41060e",
42
  "mllm_image_size_policy": "External source and prediction may differ in size; both full RGB images are sent unchanged, no crop/resize/padding/compositing. Pixel protocols are unchanged.",
43
+ "evaluation_code_note": "Frozen generation/runtime snapshot with the independently sized MLLM input compatibility fix; original source_commit/code_commit identifies generation.",
44
+ "mllm_smoke_test": {
45
+ "path": "validation/joyai_gemini16",
46
+ "model": "joyai-image-edit-plus",
47
+ "judge_model": "gemini-4-argon",
48
+ "cases": 16,
49
+ "complete": true,
50
+ "full_benchmark": false
51
+ }
52
  }
validation/joyai_gemini16/README.md ADDED
@@ -0,0 +1,12 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # JoyAI × Gemini:16-case API smoke test
2
+
3
+ Text8 + assets8; original full source + official full output at different dimensions; no image resize or crop.
4
+
5
+ | Track | Scored | IF | TA/OA | VC | LP | All four >=4 |
6
+ |---|---:|---:|---:|---:|---:|---:|
7
+ | text | 8/8 | 2.250 | 3.000 | 2.875 | 2.000 | 0/8 |
8
+ | assets | 8/8 | 2.125 | 2.375 | 2.875 | 2.125 | 0/8 |
9
+
10
+ API calls including retries:19. Model images unchanged:True.
11
+
12
+ This subset checks API connectivity and input compatibility; it is not a full benchmark ranking.
validation/joyai_gemini16/case_scores.csv ADDED
@@ -0,0 +1,17 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ track,id,instruction,status,IF,TA,OA,VC,LP,joint_qualified,IF_rationale,TA_rationale,OA_rationale,VC_rationale,LP_rationale
2
+ text,052_move,Move '0kSKUs' below 'Z1RB9 Empty' slightly down.,scored,2,2,,2,2,False,Moved the entire card up-left instead of moving the text down.,Target text became garbled and duplicated across overlapping cards.,,"Severe glyph distortion, overlapping cards, and degraded text quality.",Overlaps neighboring cards and degrades small text across the entire image.
3
+ text,094_move,Move 'Third Quarter Update' slightly down.,scored,1,2,,2,2,False,Target text was corrupted in place instead of moved down.,'Third Quarter' is severely garbled; only 'Update' remains legible.,,Malformed strokes and incoherent glyphs across the edited title.,Widespread text corruption across non-target headers and body text.
4
+ text,180_delete,Delete 'From Kitchen Chemicals to a Diagnostic Reaction'.,scored,2,2,,3,2,False,Only 'Diagnostic' was deleted; the rest of the phrase remains.,Most of the target phrase remains visible instead of being deleted.,,"Remaining text is legible, though slightly degraded in sharpness.",Widespread text corruption and glyph distortion across non-target regions.
5
+ text,315_replace,Replace 'Loss overview' with 'Loss summary'.,scored,4,4,,3,2,False,Target text 'Loss overview' replaced with 'Loss summary'.,Target wording 'Loss summary' is spelled accurately.,,Edited text integrates reasonably with minor glyph softness.,Extensive small non-target text degraded and garbled globally.
6
+ text,333_color,Change the color of 'Year' to dark red.,scored,3,4,,3,2,False,"Color changed to dark red, but text was unrequestedly enlarged.",Target text 'Year' is accurately spelled with proper case.,,Rendered text is legible but disproportionately large and bold.,Surrounding smaller text across the chart shows noticeable degradation.
7
+ text,373_delete,Delete 'Tool Stacks'.,scored,2,2,,3,2,False,Target text remains and was duplicated nearby.,Original 'Tool Stacks' was not deleted.,,"Clean background fill, though glyphs degraded globally.",Removed Framework B-D boxes and degraded surrounding text.
8
+ text,385_scale,Make 'Enforcing Event Logic in LMs' smaller.,scored,1,4,,4,2,False,Target text size was not reduced.,Target wording remains accurate and complete.,,Target text is legible with slight softness.,Severe garbling of unrelated text throughout panels.
9
+ text,396_insert,Add '(c)' beside the Model C heading.,scored,3,4,,3,2,False,Added '(c)' on the right side instead of matching left alignment.,Exact requested characters '(c)' are present.,,"Inserted text is legible, with a small stray dot nearby.",Extensive corruption and gibberish across non-target text throughout the image.
10
+ assets,asset_000_a00_scale,"Make the photo of salmon, cucumber and avocado sushi rolls near the top center smaller.",scored,1,,2,3,2,False,"Photo was not scaled smaller; instead, extra sushi rolls were added.",,Photo size unchanged and internal content altered with two extra rolls.,Added sushi rolls break out of the top photo boundary unnaturally.,Widespread text degradation and garbling across non-target regions.
11
+ assets,asset_004_a01_replace,Replace the photo of a curved terracotta staircase in the lower-right area with a photo of the same staircase beside a pale blue wall.,scored,2,,2,3,2,False,Deleted lower-right photo; replaced center photo instead.,,New staircase placed in wrong center slot.,"Clean photo integration, degraded global text clarity.",Center photo/card overwritten and small text garbled.
12
+ assets,asset_019_a01_insert,Add a photo of a tidy living room with green cushions in the empty image area in the middle-left area.,scored,2,,2,3,2,False,Expanded existing top-left photo downward instead of filling the empty testimonial slot.,,Empty middle-left slot remains blank; green cushions added to wrong upper photo.,"Expanded living room renders cleanly, though surrounding small text shows degradation.","Deleted 'Clean homes. Happy people.' heading, cleaning-bucket icon, and degraded small text."
13
+ assets,asset_138_a02_move,Move the upper-right character card showing a teal robot holding a blank card near the top center slightly to the right.,scored,2,,2,2,2,False,"Target card moved right, but experienced massive unrequested scaling and rotation.",,"Robot redrawn in wrong style, cropped at waist, missing blank card.",Oversized card poorly integrated with mismatched 2D style and heavy drop shadow.,Enlarged card occludes neighboring panels; small text degraded across entire image.
14
+ assets,asset_144_a02_move,Move the stretching woman in a purple top and orange trousers in the upper-right area slightly to the right.,scored,2,,2,3,2,False,"Woman shifted slightly right, but pose was redrawn instead of pure translation.",,Target woman completely changed pose and identity instead of preserving original asset.,Redrawn woman has slightly distorted anatomy and floating hair.,"Minor text corruptions across cards, such as 'usual' becoming 'uuual'."
15
+ assets,asset_149_a02_delete,Delete the illustration of three reusable bottles in the upper-right area.,scored,3,,3,3,3,False,Bottle illustration under 05 completely deleted.,,Target three bottles and box completely absent.,Clean background fill where bottles were.,Slight vertical cropping at top and bottom margins.
16
+ assets,asset_155_a00_replace,Replace the leftmost full-body cartoon child in a yellow jacket in the upper-left area with the same cartoon child in a blue jacket.,scored,3,,3,3,2,False,"Leftmost cartoon child replaced, but enlarged and slightly changed pose.",,"Blue jacket present, but child's proportions and character consistency changed.","Clean boundaries, though style and scale differ from adjacent characters.",Global text degradation and garbled characters across all non-target cards.
17
+ assets,asset_171_a03_insert,Add an illustration of a child in a blue cap sitting beside a gray cat in the empty image area in the lower-right area.,scored,2,,3,3,2,False,"Inserted asset in lower-right, but replaced existing illustrations instead of empty space.",,"Child in blue cap and gray cat match attributes, but wrong slot.","Clean illustration style, though child overlaps text box and clips word.","Deleted woman, child, dog, and tent; occluded 'child' text in bottom box."
validation/joyai_gemini16/selection.json ADDED
@@ -0,0 +1,222 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model": "joyai-image-edit-plus",
3
+ "judge_model": "gemini-4-argon",
4
+ "sampling": "operation-balanced and source-diverse,seed42;8 tasks pertrack; no outcome-based selection",
5
+ "hf_result_revision_used": "94c6854368150b8690f9f4d416079f474caf8384",
6
+ "evaluation_code_commit": "2dc67b321c948d1976dc7bf5c00910e7cd41060e",
7
+ "cases": {
8
+ "text": [
9
+ {
10
+ "id": "315_replace",
11
+ "operation": "replace",
12
+ "source_id": "315",
13
+ "source_size": [
14
+ 1600,
15
+ 900
16
+ ],
17
+ "prediction_size": [
18
+ 1344,
19
+ 768
20
+ ]
21
+ },
22
+ {
23
+ "id": "180_delete",
24
+ "operation": "delete",
25
+ "source_id": "180",
26
+ "source_size": [
27
+ 1600,
28
+ 900
29
+ ],
30
+ "prediction_size": [
31
+ 1344,
32
+ 768
33
+ ]
34
+ },
35
+ {
36
+ "id": "373_delete",
37
+ "operation": "delete",
38
+ "source_id": "373",
39
+ "source_size": [
40
+ 1600,
41
+ 900
42
+ ],
43
+ "prediction_size": [
44
+ 1344,
45
+ 768
46
+ ]
47
+ },
48
+ {
49
+ "id": "396_insert",
50
+ "operation": "insert",
51
+ "source_id": "396",
52
+ "source_size": [
53
+ 1400,
54
+ 2440
55
+ ],
56
+ "prediction_size": [
57
+ 768,
58
+ 1344
59
+ ]
60
+ },
61
+ {
62
+ "id": "052_move",
63
+ "operation": "move",
64
+ "source_id": "052",
65
+ "source_size": [
66
+ 1600,
67
+ 1200
68
+ ],
69
+ "prediction_size": [
70
+ 1152,
71
+ 832
72
+ ]
73
+ },
74
+ {
75
+ "id": "385_scale",
76
+ "operation": "scale",
77
+ "source_id": "385",
78
+ "source_size": [
79
+ 1600,
80
+ 1410
81
+ ],
82
+ "prediction_size": [
83
+ 1088,
84
+ 960
85
+ ]
86
+ },
87
+ {
88
+ "id": "333_color",
89
+ "operation": "color",
90
+ "source_id": "333",
91
+ "source_size": [
92
+ 1600,
93
+ 900
94
+ ],
95
+ "prediction_size": [
96
+ 1344,
97
+ 768
98
+ ]
99
+ },
100
+ {
101
+ "id": "094_move",
102
+ "operation": "move",
103
+ "source_id": "094",
104
+ "source_size": [
105
+ 1600,
106
+ 900
107
+ ],
108
+ "prediction_size": [
109
+ 1344,
110
+ 768
111
+ ]
112
+ }
113
+ ],
114
+ "assets": [
115
+ {
116
+ "id": "asset_144_a02_move",
117
+ "operation": "move",
118
+ "source_id": "144",
119
+ "source_size": [
120
+ 1200,
121
+ 1600
122
+ ],
123
+ "prediction_size": [
124
+ 896,
125
+ 1152
126
+ ]
127
+ },
128
+ {
129
+ "id": "asset_149_a02_delete",
130
+ "operation": "delete",
131
+ "source_id": "149",
132
+ "source_size": [
133
+ 1200,
134
+ 1800
135
+ ],
136
+ "prediction_size": [
137
+ 832,
138
+ 1216
139
+ ]
140
+ },
141
+ {
142
+ "id": "asset_155_a00_replace",
143
+ "operation": "replace",
144
+ "source_id": "155",
145
+ "source_size": [
146
+ 1200,
147
+ 2133
148
+ ],
149
+ "prediction_size": [
150
+ 768,
151
+ 1344
152
+ ]
153
+ },
154
+ {
155
+ "id": "asset_171_a03_insert",
156
+ "operation": "insert",
157
+ "source_id": "171",
158
+ "source_size": [
159
+ 1200,
160
+ 1800
161
+ ],
162
+ "prediction_size": [
163
+ 832,
164
+ 1216
165
+ ]
166
+ },
167
+ {
168
+ "id": "asset_138_a02_move",
169
+ "operation": "move",
170
+ "source_id": "138",
171
+ "source_size": [
172
+ 1440,
173
+ 2560
174
+ ],
175
+ "prediction_size": [
176
+ 768,
177
+ 1344
178
+ ]
179
+ },
180
+ {
181
+ "id": "asset_004_a01_replace",
182
+ "operation": "replace",
183
+ "source_id": "004",
184
+ "source_size": [
185
+ 1600,
186
+ 1200
187
+ ],
188
+ "prediction_size": [
189
+ 1152,
190
+ 832
191
+ ]
192
+ },
193
+ {
194
+ "id": "asset_019_a01_insert",
195
+ "operation": "insert",
196
+ "source_id": "019",
197
+ "source_size": [
198
+ 1600,
199
+ 1200
200
+ ],
201
+ "prediction_size": [
202
+ 1152,
203
+ 832
204
+ ]
205
+ },
206
+ {
207
+ "id": "asset_000_a00_scale",
208
+ "operation": "scale",
209
+ "source_id": "000",
210
+ "source_size": [
211
+ 1600,
212
+ 1200
213
+ ],
214
+ "prediction_size": [
215
+ 1152,
216
+ 832
217
+ ]
218
+ }
219
+ ]
220
+ },
221
+ "input_policy": "Full original source and full official JoyAI output at their own dimensions; no resize/crop/padding/compositing."
222
+ }
validation/joyai_gemini16/summary.json ADDED
@@ -0,0 +1,638 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "status": "complete",
3
+ "model": "joyai-image-edit-plus",
4
+ "judge_model": "gemini-4-argon",
5
+ "cases": 16,
6
+ "api_calls": 19,
7
+ "tracks": {
8
+ "text": {
9
+ "expected": 8,
10
+ "scored": 8,
11
+ "pending": 0,
12
+ "errors": 0,
13
+ "criteria": {
14
+ "IF": {
15
+ "mean": 2.25,
16
+ "scored": 8,
17
+ "expected": 8,
18
+ "coverage": 1.0
19
+ },
20
+ "TA": {
21
+ "mean": 3.0,
22
+ "scored": 8,
23
+ "expected": 8,
24
+ "coverage": 1.0
25
+ },
26
+ "VC": {
27
+ "mean": 2.875,
28
+ "scored": 8,
29
+ "expected": 8,
30
+ "coverage": 1.0
31
+ },
32
+ "LP": {
33
+ "mean": 2.0,
34
+ "scored": 8,
35
+ "expected": 8,
36
+ "coverage": 1.0
37
+ },
38
+ "SE": {
39
+ "mean": null,
40
+ "scored": 0,
41
+ "expected": 0,
42
+ "coverage": null
43
+ }
44
+ },
45
+ "joint_qualified_count": 0,
46
+ "joint_qualified_rate": 0.0
47
+ },
48
+ "assets": {
49
+ "expected": 8,
50
+ "valid_outputs": 8,
51
+ "model_failures": 0,
52
+ "scored": 8,
53
+ "pending": 0,
54
+ "errors": 0,
55
+ "formal_complete": true,
56
+ "valid_output_coverage": 1.0,
57
+ "criteria": {
58
+ "IF": {
59
+ "mean": 2.125,
60
+ "scored": 8,
61
+ "expected": 8,
62
+ "judge_expected": 8,
63
+ "coverage": 1.0,
64
+ "judge_coverage": 1.0,
65
+ "mean_denominator": "valid model outputs only"
66
+ },
67
+ "OA": {
68
+ "mean": 2.375,
69
+ "scored": 8,
70
+ "expected": 8,
71
+ "judge_expected": 8,
72
+ "coverage": 1.0,
73
+ "judge_coverage": 1.0,
74
+ "mean_denominator": "valid model outputs only"
75
+ },
76
+ "VC": {
77
+ "mean": 2.875,
78
+ "scored": 8,
79
+ "expected": 8,
80
+ "judge_expected": 8,
81
+ "coverage": 1.0,
82
+ "judge_coverage": 1.0,
83
+ "mean_denominator": "valid model outputs only"
84
+ },
85
+ "LP": {
86
+ "mean": 2.125,
87
+ "scored": 8,
88
+ "expected": 8,
89
+ "judge_expected": 8,
90
+ "coverage": 1.0,
91
+ "judge_coverage": 1.0,
92
+ "mean_denominator": "valid model outputs only"
93
+ }
94
+ },
95
+ "four_dimension_mean": 2.375,
96
+ "joint_qualification": {
97
+ "threshold_each": 4,
98
+ "qualified": 0,
99
+ "expected": 8,
100
+ "rate": 0.0,
101
+ "model_failures_count_as_unqualified": true
102
+ },
103
+ "consistency_review": {
104
+ "cases": 0,
105
+ "case_ids": [],
106
+ "scores_adjusted": false
107
+ },
108
+ "by_operation": {
109
+ "delete": {
110
+ "expected": 1,
111
+ "valid_outputs": 1,
112
+ "model_failures": 0,
113
+ "scored": 1,
114
+ "pending": 0,
115
+ "errors": 0,
116
+ "formal_complete": true,
117
+ "valid_output_coverage": 1.0,
118
+ "criteria": {
119
+ "IF": {
120
+ "mean": 3.0,
121
+ "scored": 1,
122
+ "expected": 1,
123
+ "judge_expected": 1,
124
+ "coverage": 1.0,
125
+ "judge_coverage": 1.0,
126
+ "mean_denominator": "valid model outputs only"
127
+ },
128
+ "OA": {
129
+ "mean": 3.0,
130
+ "scored": 1,
131
+ "expected": 1,
132
+ "judge_expected": 1,
133
+ "coverage": 1.0,
134
+ "judge_coverage": 1.0,
135
+ "mean_denominator": "valid model outputs only"
136
+ },
137
+ "VC": {
138
+ "mean": 3.0,
139
+ "scored": 1,
140
+ "expected": 1,
141
+ "judge_expected": 1,
142
+ "coverage": 1.0,
143
+ "judge_coverage": 1.0,
144
+ "mean_denominator": "valid model outputs only"
145
+ },
146
+ "LP": {
147
+ "mean": 3.0,
148
+ "scored": 1,
149
+ "expected": 1,
150
+ "judge_expected": 1,
151
+ "coverage": 1.0,
152
+ "judge_coverage": 1.0,
153
+ "mean_denominator": "valid model outputs only"
154
+ }
155
+ },
156
+ "four_dimension_mean": 3.0,
157
+ "joint_qualification": {
158
+ "threshold_each": 4,
159
+ "qualified": 0,
160
+ "expected": 1,
161
+ "rate": 0.0,
162
+ "model_failures_count_as_unqualified": true
163
+ }
164
+ },
165
+ "insert": {
166
+ "expected": 2,
167
+ "valid_outputs": 2,
168
+ "model_failures": 0,
169
+ "scored": 2,
170
+ "pending": 0,
171
+ "errors": 0,
172
+ "formal_complete": true,
173
+ "valid_output_coverage": 1.0,
174
+ "criteria": {
175
+ "IF": {
176
+ "mean": 2.0,
177
+ "scored": 2,
178
+ "expected": 2,
179
+ "judge_expected": 2,
180
+ "coverage": 1.0,
181
+ "judge_coverage": 1.0,
182
+ "mean_denominator": "valid model outputs only"
183
+ },
184
+ "OA": {
185
+ "mean": 2.5,
186
+ "scored": 2,
187
+ "expected": 2,
188
+ "judge_expected": 2,
189
+ "coverage": 1.0,
190
+ "judge_coverage": 1.0,
191
+ "mean_denominator": "valid model outputs only"
192
+ },
193
+ "VC": {
194
+ "mean": 3.0,
195
+ "scored": 2,
196
+ "expected": 2,
197
+ "judge_expected": 2,
198
+ "coverage": 1.0,
199
+ "judge_coverage": 1.0,
200
+ "mean_denominator": "valid model outputs only"
201
+ },
202
+ "LP": {
203
+ "mean": 2.0,
204
+ "scored": 2,
205
+ "expected": 2,
206
+ "judge_expected": 2,
207
+ "coverage": 1.0,
208
+ "judge_coverage": 1.0,
209
+ "mean_denominator": "valid model outputs only"
210
+ }
211
+ },
212
+ "four_dimension_mean": 2.375,
213
+ "joint_qualification": {
214
+ "threshold_each": 4,
215
+ "qualified": 0,
216
+ "expected": 2,
217
+ "rate": 0.0,
218
+ "model_failures_count_as_unqualified": true
219
+ }
220
+ },
221
+ "replace": {
222
+ "expected": 2,
223
+ "valid_outputs": 2,
224
+ "model_failures": 0,
225
+ "scored": 2,
226
+ "pending": 0,
227
+ "errors": 0,
228
+ "formal_complete": true,
229
+ "valid_output_coverage": 1.0,
230
+ "criteria": {
231
+ "IF": {
232
+ "mean": 2.5,
233
+ "scored": 2,
234
+ "expected": 2,
235
+ "judge_expected": 2,
236
+ "coverage": 1.0,
237
+ "judge_coverage": 1.0,
238
+ "mean_denominator": "valid model outputs only"
239
+ },
240
+ "OA": {
241
+ "mean": 2.5,
242
+ "scored": 2,
243
+ "expected": 2,
244
+ "judge_expected": 2,
245
+ "coverage": 1.0,
246
+ "judge_coverage": 1.0,
247
+ "mean_denominator": "valid model outputs only"
248
+ },
249
+ "VC": {
250
+ "mean": 3.0,
251
+ "scored": 2,
252
+ "expected": 2,
253
+ "judge_expected": 2,
254
+ "coverage": 1.0,
255
+ "judge_coverage": 1.0,
256
+ "mean_denominator": "valid model outputs only"
257
+ },
258
+ "LP": {
259
+ "mean": 2.0,
260
+ "scored": 2,
261
+ "expected": 2,
262
+ "judge_expected": 2,
263
+ "coverage": 1.0,
264
+ "judge_coverage": 1.0,
265
+ "mean_denominator": "valid model outputs only"
266
+ }
267
+ },
268
+ "four_dimension_mean": 2.5,
269
+ "joint_qualification": {
270
+ "threshold_each": 4,
271
+ "qualified": 0,
272
+ "expected": 2,
273
+ "rate": 0.0,
274
+ "model_failures_count_as_unqualified": true
275
+ }
276
+ },
277
+ "move": {
278
+ "expected": 2,
279
+ "valid_outputs": 2,
280
+ "model_failures": 0,
281
+ "scored": 2,
282
+ "pending": 0,
283
+ "errors": 0,
284
+ "formal_complete": true,
285
+ "valid_output_coverage": 1.0,
286
+ "criteria": {
287
+ "IF": {
288
+ "mean": 2.0,
289
+ "scored": 2,
290
+ "expected": 2,
291
+ "judge_expected": 2,
292
+ "coverage": 1.0,
293
+ "judge_coverage": 1.0,
294
+ "mean_denominator": "valid model outputs only"
295
+ },
296
+ "OA": {
297
+ "mean": 2.0,
298
+ "scored": 2,
299
+ "expected": 2,
300
+ "judge_expected": 2,
301
+ "coverage": 1.0,
302
+ "judge_coverage": 1.0,
303
+ "mean_denominator": "valid model outputs only"
304
+ },
305
+ "VC": {
306
+ "mean": 2.5,
307
+ "scored": 2,
308
+ "expected": 2,
309
+ "judge_expected": 2,
310
+ "coverage": 1.0,
311
+ "judge_coverage": 1.0,
312
+ "mean_denominator": "valid model outputs only"
313
+ },
314
+ "LP": {
315
+ "mean": 2.0,
316
+ "scored": 2,
317
+ "expected": 2,
318
+ "judge_expected": 2,
319
+ "coverage": 1.0,
320
+ "judge_coverage": 1.0,
321
+ "mean_denominator": "valid model outputs only"
322
+ }
323
+ },
324
+ "four_dimension_mean": 2.125,
325
+ "joint_qualification": {
326
+ "threshold_each": 4,
327
+ "qualified": 0,
328
+ "expected": 2,
329
+ "rate": 0.0,
330
+ "model_failures_count_as_unqualified": true
331
+ }
332
+ },
333
+ "scale": {
334
+ "expected": 1,
335
+ "valid_outputs": 1,
336
+ "model_failures": 0,
337
+ "scored": 1,
338
+ "pending": 0,
339
+ "errors": 0,
340
+ "formal_complete": true,
341
+ "valid_output_coverage": 1.0,
342
+ "criteria": {
343
+ "IF": {
344
+ "mean": 1.0,
345
+ "scored": 1,
346
+ "expected": 1,
347
+ "judge_expected": 1,
348
+ "coverage": 1.0,
349
+ "judge_coverage": 1.0,
350
+ "mean_denominator": "valid model outputs only"
351
+ },
352
+ "OA": {
353
+ "mean": 2.0,
354
+ "scored": 1,
355
+ "expected": 1,
356
+ "judge_expected": 1,
357
+ "coverage": 1.0,
358
+ "judge_coverage": 1.0,
359
+ "mean_denominator": "valid model outputs only"
360
+ },
361
+ "VC": {
362
+ "mean": 3.0,
363
+ "scored": 1,
364
+ "expected": 1,
365
+ "judge_expected": 1,
366
+ "coverage": 1.0,
367
+ "judge_coverage": 1.0,
368
+ "mean_denominator": "valid model outputs only"
369
+ },
370
+ "LP": {
371
+ "mean": 2.0,
372
+ "scored": 1,
373
+ "expected": 1,
374
+ "judge_expected": 1,
375
+ "coverage": 1.0,
376
+ "judge_coverage": 1.0,
377
+ "mean_denominator": "valid model outputs only"
378
+ }
379
+ },
380
+ "four_dimension_mean": 2.0,
381
+ "joint_qualification": {
382
+ "threshold_each": 4,
383
+ "qualified": 0,
384
+ "expected": 1,
385
+ "rate": 0.0,
386
+ "model_failures_count_as_unqualified": true
387
+ }
388
+ }
389
+ },
390
+ "joint_qualified_count": 0,
391
+ "joint_qualified_rate": 0.0
392
+ }
393
+ },
394
+ "inputs_unchanged": true,
395
+ "input_policy": "Full original source and full official JoyAI output at their own dimensions; no resize/crop/padding/compositing.",
396
+ "representative_only": true,
397
+ "full_benchmark_score": false,
398
+ "selection": {
399
+ "model": "joyai-image-edit-plus",
400
+ "judge_model": "gemini-4-argon",
401
+ "sampling": "operation-balanced and source-diverse,seed42;8 tasks pertrack; no outcome-based selection",
402
+ "hf_result_revision_used": "94c6854368150b8690f9f4d416079f474caf8384",
403
+ "evaluation_code_commit": "2dc67b321c948d1976dc7bf5c00910e7cd41060e",
404
+ "cases": {
405
+ "text": [
406
+ {
407
+ "id": "315_replace",
408
+ "operation": "replace",
409
+ "source_id": "315",
410
+ "source_size": [
411
+ 1600,
412
+ 900
413
+ ],
414
+ "prediction_size": [
415
+ 1344,
416
+ 768
417
+ ]
418
+ },
419
+ {
420
+ "id": "180_delete",
421
+ "operation": "delete",
422
+ "source_id": "180",
423
+ "source_size": [
424
+ 1600,
425
+ 900
426
+ ],
427
+ "prediction_size": [
428
+ 1344,
429
+ 768
430
+ ]
431
+ },
432
+ {
433
+ "id": "373_delete",
434
+ "operation": "delete",
435
+ "source_id": "373",
436
+ "source_size": [
437
+ 1600,
438
+ 900
439
+ ],
440
+ "prediction_size": [
441
+ 1344,
442
+ 768
443
+ ]
444
+ },
445
+ {
446
+ "id": "396_insert",
447
+ "operation": "insert",
448
+ "source_id": "396",
449
+ "source_size": [
450
+ 1400,
451
+ 2440
452
+ ],
453
+ "prediction_size": [
454
+ 768,
455
+ 1344
456
+ ]
457
+ },
458
+ {
459
+ "id": "052_move",
460
+ "operation": "move",
461
+ "source_id": "052",
462
+ "source_size": [
463
+ 1600,
464
+ 1200
465
+ ],
466
+ "prediction_size": [
467
+ 1152,
468
+ 832
469
+ ]
470
+ },
471
+ {
472
+ "id": "385_scale",
473
+ "operation": "scale",
474
+ "source_id": "385",
475
+ "source_size": [
476
+ 1600,
477
+ 1410
478
+ ],
479
+ "prediction_size": [
480
+ 1088,
481
+ 960
482
+ ]
483
+ },
484
+ {
485
+ "id": "333_color",
486
+ "operation": "color",
487
+ "source_id": "333",
488
+ "source_size": [
489
+ 1600,
490
+ 900
491
+ ],
492
+ "prediction_size": [
493
+ 1344,
494
+ 768
495
+ ]
496
+ },
497
+ {
498
+ "id": "094_move",
499
+ "operation": "move",
500
+ "source_id": "094",
501
+ "source_size": [
502
+ 1600,
503
+ 900
504
+ ],
505
+ "prediction_size": [
506
+ 1344,
507
+ 768
508
+ ]
509
+ }
510
+ ],
511
+ "assets": [
512
+ {
513
+ "id": "asset_144_a02_move",
514
+ "operation": "move",
515
+ "source_id": "144",
516
+ "source_size": [
517
+ 1200,
518
+ 1600
519
+ ],
520
+ "prediction_size": [
521
+ 896,
522
+ 1152
523
+ ]
524
+ },
525
+ {
526
+ "id": "asset_149_a02_delete",
527
+ "operation": "delete",
528
+ "source_id": "149",
529
+ "source_size": [
530
+ 1200,
531
+ 1800
532
+ ],
533
+ "prediction_size": [
534
+ 832,
535
+ 1216
536
+ ]
537
+ },
538
+ {
539
+ "id": "asset_155_a00_replace",
540
+ "operation": "replace",
541
+ "source_id": "155",
542
+ "source_size": [
543
+ 1200,
544
+ 2133
545
+ ],
546
+ "prediction_size": [
547
+ 768,
548
+ 1344
549
+ ]
550
+ },
551
+ {
552
+ "id": "asset_171_a03_insert",
553
+ "operation": "insert",
554
+ "source_id": "171",
555
+ "source_size": [
556
+ 1200,
557
+ 1800
558
+ ],
559
+ "prediction_size": [
560
+ 832,
561
+ 1216
562
+ ]
563
+ },
564
+ {
565
+ "id": "asset_138_a02_move",
566
+ "operation": "move",
567
+ "source_id": "138",
568
+ "source_size": [
569
+ 1440,
570
+ 2560
571
+ ],
572
+ "prediction_size": [
573
+ 768,
574
+ 1344
575
+ ]
576
+ },
577
+ {
578
+ "id": "asset_004_a01_replace",
579
+ "operation": "replace",
580
+ "source_id": "004",
581
+ "source_size": [
582
+ 1600,
583
+ 1200
584
+ ],
585
+ "prediction_size": [
586
+ 1152,
587
+ 832
588
+ ]
589
+ },
590
+ {
591
+ "id": "asset_019_a01_insert",
592
+ "operation": "insert",
593
+ "source_id": "019",
594
+ "source_size": [
595
+ 1600,
596
+ 1200
597
+ ],
598
+ "prediction_size": [
599
+ 1152,
600
+ 832
601
+ ]
602
+ },
603
+ {
604
+ "id": "asset_000_a00_scale",
605
+ "operation": "scale",
606
+ "source_id": "000",
607
+ "source_size": [
608
+ 1600,
609
+ 1200
610
+ ],
611
+ "prediction_size": [
612
+ 1152,
613
+ 832
614
+ ]
615
+ }
616
+ ]
617
+ },
618
+ "input_policy": "Full original source and full official JoyAI output at their own dimensions; no resize/crop/padding/compositing."
619
+ },
620
+ "job_id": "12087",
621
+ "completed_at": "2026-11-08T15:03:36.047613+00:00",
622
+ "request_parameters": {
623
+ "temperature": 0,
624
+ "max_tokens": 500
625
+ },
626
+ "requests_per_case": 1,
627
+ "request_status_counts": {
628
+ "scored": 16,
629
+ "error": 3
630
+ },
631
+ "request_error_counts": {
632
+ "Judge returned invalid JSON": 1,
633
+ "API connection/timeout failure": 2
634
+ },
635
+ "returned_model_counts": {
636
+ "gemini-4-argon": 16
637
+ }
638
+ }