TextSculpt-Bench: leaderboard
TextSculpt-Bench (2026) tests image editing models on 800 scene-text edits: adding, removing and replacing text, and mixed edits, in photos and generated posters. The score averages OCR-based text accuracy, GPT-5.2-judged visual quality and SSIM background preservation, each scaled to 100 (SSIM can in principle fall below 0); higher is better.
Metric: Mean of text accuracy and visual quality (each 0 to 100) and background preservation (SSIM x 100, -100 to 100), so -33.3 to 100, over 800 TextSculpt-Bench text-editing tasks (200 each of addition, removal, replacement and hybrid editing) on Pexels photos and generated posters, default model settings; higher is better. Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 9 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Seedream 4.5 | 72 | |
| 2 | Gemini 2.5 Flash Image | 68 | |
| 3 | FireRed-Image-Edit-1.0 | 68 | |
| 4 | Qwen-Image-Edit-2511 | 65 | |
| 5 | Seedream 4.0 | 59 | |
| 6 | LongCat-Image-Edit | 58 | |
| 7 | Step1X-Edit | 55 | |
| 8 | Bagel | 43 | |
| 9 | OmniGen2 | 39 | #4579 |
Interactive version: theaggregate.ai/benchmark?slug=textsculpt-bench · How It Works · Data refreshed daily, snapshot 2026-11-21.