MobileForge - Visual Fidelity: leaderboard
MobileForge (Harbin Institute of Technology and others, 2026) evaluates project-level mobile app generation from the screenshots of 29 real apps. This board is point-wise visual fidelity: a Gemini 2.5 Pro judge rates each generated screen against its reference design on a 1 to 5 rubric; higher is better.
Metric: Point-wise visual fidelity (1-5): mean four-dimension rubric score a Gemini 2.5 Pro vision-language judge gives each generated screen against its reference design; 29 real mobile apps given as 309 screenshots with page-relationship annotations, each turned into a React, TypeScript and Tailwind project by one tool-using agent in a fixed harness (up to 50 iterations, build errors fed back); one run per app; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.6 | 2.99 | #102 |
| 2 | GPT-5 | 2.64 | #167 |
| 3 | Gemini 2.5 Pro | 2.56 | #266 |
| 4 | Claude Haiku 4.5 | 2.37 | #568 |
| 5 | Gemini 2.5 Flash | 2 | #453 |
| 6 | GPT-5 Mini | 1.6 | #348 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=mobileforge-visual-fidelity · How It Works · Data refreshed daily, snapshot 2026-11-13.