File size: 4,477 Bytes
47f13c4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
---
pretty_name: Scry Design Diff Eval
license: cc-by-4.0
task_categories:
- image-to-text
- object-detection
tags:
- ui
- mobile
- visual-regression
- benchmark
- vlm
- screenshots
size_categories:
- n<1K
configs:
- config_name: default
  data_files:
  - split: test
    path: data/test-*
---

# Scry Design Diff Eval

**Measuring VLMs as Mobile UI Regression Reviewers**

Scry Design Diff Eval is a benchmark built to evaluate vision-language models on
mobile UI diff review. Each example pairs a reference mobile screenshot
(`image_a`) with a generated implementation screenshot (`image_b`) and carries
human-drawn selection boxes plus explicit defect tags. A model must return a
structured list of UI defects — tags and normalized boxes — not a prose
description.

📄 **Paper**: https://blog.scrymore.com/

## Dataset composition

| Quantity | Count |
|---|---:|
| Eval pairs | 311 |
| Visual-diff pairs | 234 |
| No-tagged controls | 77 |
| Counted tagged issues | 557 |

Issue-density buckets (the `split` column):

| Bucket | Definition | Pairs |
|---|---|---:|
| `single_issue` | 1 tagged issue | 107 |
| `multi_issue` | 2-3 tagged issues | 80 |
| `dense_issue` | 4 or more tagged issues | 47 |
| `no_diff` | 0 scored tagged issues | 77 |

The most common defect tags are `Icon/Nav`, `Color/Background`,
`Spacing/Layout`, `Shape/Size`, `Missing Content`, and `Typography`.

## Schema

| Column | Type | Description |
|---|---|---|
| `id` | string | Pair id, e.g. `amazon-shopping__801` |
| `app_name` | string | Source app and capture batch |
| `app_slug` | string | Normalized app identifier |
| `screen_index` | int | Screen number within the app capture |
| `split` | string | Issue-density bucket (see above) |
| `task_type` | string | `visual_diff` or `no_diff` |
| `n_issues` | int | Number of scored tagged issues |
| `issue_labels` | list[string] | Union of defect tags on this pair |
| `ground_truth_issues` | string | JSON array of issues: `{issue_id, labels, box_a, box_b, note?, created_at}` with boxes normalized to `{x, y, w, h}` in [0, 1] |
| `image_a` | image | Reference screenshot |
| `image_b` | image | Generated implementation screenshot |

Parse `ground_truth_issues` with `json.loads`. A box may be present on side A,
side B, or both.

## Task

Given `image_a` and `image_b`, return:

```json
{
  "issues": [
    {
      "labels": ["Icon/Nav"],
      "note": "The bottom navigation icon differs from the reference.",
      "box_a": {"x": 0.10, "y": 0.90, "w": 0.12, "h": 0.07},
      "box_b": {"x": 0.10, "y": 0.90, "w": 0.12, "h": 0.07},
      "confidence": 0.80
    }
  ]
}
```

## Scoring

The primary metric is **known-issue recall**. A model issue matches a human
issue when they share at least one explicit defect tag AND the model box
overlaps the human box on the same image side with IoU >= 0.10. Matching is
one-to-one. The judge is deterministic — no LLM judging.

The protocol is recall-first because human annotations are known positives, not
exhaustive negatives: extra model findings may be valid and are reported
diagnostically (precision, no-tagged flag rate) rather than reducing the
primary score.

## Baseline results (full 311-pair set)

| Model | Known-Issue Recall | Diagnostic Precision | Issue F1 |
|---|---:|---:|---:|
| Kimi K2.7 Code + Together recovery | 38.2% | 15.9% | 22.5% |
| Gemini 3.5 Flash | 37.5% | 20.2% | 26.3% |
| Codex GPT-5.5 xhigh | 37.3% | 13.6% | 19.9% |
| MiniMax M3 | 21.9% | 11.5% | 15.0% |
| Gemma 4 26B A4B | 20.3% | 12.4% | 15.4% |
| Gemma 4 31B | 17.8% | 13.3% | 15.2% |

See the paper for pilot results, density and category breakdowns, and static
controls.

## Caveats

- **No-tagged controls are a proxy, not a guarantee**: pairs with zero scored
  tagged issues are used as controls but were not exhaustively audited as
  defect-free.
- **Annotations are known positives**: a model can find a real defect the
  annotators did not tag. Treat precision and no-diff specificity as
  operational diagnostics.
- **Screenshots**: reference images are captures of real mobile apps, included
  for research and evaluation purposes; all app content remains the property
  of its respective owners. The annotations (boxes, tags, metadata) are
  released under CC BY 4.0.

## Citation

```bibtex
@misc{scrydesigndiffeval2026,
  title={Scry Design Diff Eval: Measuring VLMs as Mobile UI Regression Reviewers},
  author={Pinnock, Ejiro},
  year={2026},
  url={https://blog.scrymore.com/}
}
```