Image benchmarks
Every benchmark the Image index tracks, with each core benchmark's share of the composite and how many models it covers.
25 benchmarks tracked.
- Artificial Analysis Image Arena, Editing (AA-Image-Editing v2.0)
- Artificial Analysis Image Arena, Text to Image (AA-Image-T2I v2.0)
- AA T2I capability: Text Rendering Elo
- Arena-T2I-Hard faithfulness (dependency-aware checklist)
- BudgetPixel AI Image Arena, overall
- CVTG-2K (complex visual text generation)
- DPG-Bench (dense prompt graph, VQA judged)
- GEdit-Bench-EN (Step1X-Edit; G_O / Q_O overall)
- GenEval (object-focused compositional alignment)
- GenEval 2 (Soft-TIFA)
- HPSv3 learned preference score
- ImgEdit-Bench
- KRIS-Bench (knowledge/reasoning edits)
- LMArena (Arena) Image Edit, single-image overall
- LMArena Image Edit, multi-image category
- LMArena (Arena) Text-to-Image, overall
- LMArena Text-to-Image, text_rendering category
- LongText-Bench (long text rendering, EN/ZH)
- OneIG-Bench (alignment, text, reasoning, style, diversity)
- Rapidata Benchmark.ai image benchmark (Elo-style, human judged)
- T2I-CompBench++
- T2I-ReasonBench
- TIIF-Bench (instruction following, short/long prompt)
- UniGenBench++
- WISE (world-knowledge T2I)