TECCI - Image Consistency: leaderboard
Instruction-guided image editing on TECCI (Agrawal et al., 2026), newly collected images in seven categories picked for hard edits: the share of 1,315 sampled edits whose mean rating from five trained raters reaches 4.5 of 5 for leaving the rest of the image unchanged. Percent; higher is better.
Metric: Image-consistency (minimal edit) human success rate (%; five trained raters score each edit 1-5 and a criterion succeeds when the mean is at least 4.5; 1,315 sampled pairs: 265 of the Instruction Rich Challenge Set and 1,050 of the Gemini Generated Instruction Set). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Nano Banana Pro | 42.1 | |
| 2 | Grok Imagine Pro | 39.2 | |
| 3 | Seedream 5.0 Lite | 38.4 | |
| 4 | Nano Banana 2 (Gemini 3.1 Flash Image Preview) | 35.3 | |
| 5 | GPT Image 1.5 | 9.6 |
Interactive version: theaggregate.ai/benchmark?slug=tecci-image-consistency · How It Works · Data refreshed daily, snapshot 2026-11-12.