ColorBench - Color Robustness: leaderboard
Share of 493 colour-invariant image questions a model answers correctly on the original image and on all nine hue-shifted recolourings (10 images per question). Measures stability under colour change, not colour knowledge as the headline board does.
Metric: Color Robustness (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 31 models tracked. Updated 17 Nov 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | InternVL2.5-78B | 86.2 | #1036 |
| 2 | InternVL2.5-38B | 84.6 | #1366 |
| 3 | InternVL2.5-26B | 83 | #2191 |
| 4 | llava-onevision-qwen2-72B-ov-sft | 80.3 | #1957 |
| 5 | InternVL2-40B | 78.7 | #2094 |
| 6 | Qwen 2.5 VL 7B Instruct | 74.4 | #1893 |
| 7 | InternVL2-26B | 74 | #2639 |
| 8 | llava-onevision-qwen2-7B-ov | 74 | #3252 |
| 9 | Gemini 2.0 Flash | 70.7 | #723 |
| 10 | InternVL2.5-8B | 69.8 | #2401 |
| 11 | InternVL2-Llama3-76B | 68.6 | #2021 |
| 12 | Cambrian-34B | 67.7 | #5007 |
| 13 | eagle-x5-34B-plus | 67.1 | |
| 14 | LLaVA-NeXT-72B | 66.5 | #3206 |
| 15 | eagle-x4-13B-plus | 66.3 |
No result here: #1 Claude Fable 5.2, #2 GPT-6.2 Astra, #3 Gemini 4 Carbon.
Interactive version: theaggregate.ai/benchmark?slug=colorbench-color-robustness · How It Works · Data refreshed daily, snapshot 2026-11-17.