EvalAlign: Supervised Fine-Tuning Multimodal LLMs with Human-Aligned Data for Evaluating Text-to-Image Models
[Arxiv] [Code] [Model] [Dataset] [Page] [Leaderboard] [BibTeX]
| Model | Image Faithfulness (Human) | Image Faithfulness (EvalAlign) | Text-image Alignment (Human) | Text-image Alignment (EvalAlign) |
|---|---|---|---|---|
| PixArt-XL-2-1024-MS | 2.2848 | 1.6415 | 5.1100 | 5.3100 |
| Dreamlike Photoreal v2.0 | 2.0070 | 1.4522 | 4.5600 | 4.9800 |
| SDXL Refiner v1.0 | 1.9229 | 1.6072 | 5.2100 | 5.4000 |
| SDXL v1.0 | 1.8136 | 1.4675 | 5.0300 | 5.3500 |
| Wuerstchen | 1.7837 | 1.4279 | 4.8700 | 5.1700 |
| LCM SDXL | 1.6910 | 1.3391 | 5.1800 | 5.3300 |
| Openjourney | 1.6667 | 1.1750 | 4.8300 | 4.9200 |
| Safe SD MAX | 1.6491 | 1.2175 | 4.3100 | 4.5900 |
| LCM LoRA SDXL | 1.6387 | 1.3833 | 5.0600 | 5.2700 |
| Safe SD STRONG | 1.6308 | 1.1466 | 4.6000 | 4.8300 |
| Safe SD MEDIUM | 1.6275 | 1.1298 | 4.4000 | 4.5600 |
| Safe SD WEAK | 1.6078 | 1.1188 | 4.5300 | 4.7100 |
| SD v2.1 | 1.5524 | 1.1094 | 4.8000 | 5.0700 |
| SD v2.0 | 1.5277 | 1.1300 | 4.6400 | 5.0100 |
| Openjourney v2 | 1.5000 | 0.9956 | 4.1500 | 4.6500 |
| Redshift Diffusion | 1.4733 | 1.1382 | 4.3500 | 4.6700 |
| Dreamlike Diffusion v1.0 | 1.4652 | 1.2052 | 4.6600 | 5.1500 |
| SD v1.5 | 1.4417 | 1.1362 | 4.4500 | 4.9000 |
| IF-I-XL v1.0 | 1.3808 | 0.9221 | 5.4500 | 5.5300 |
| SD v1.4 | 1.3592 | 0.9511 | 4.5200 | 4.7600 |
| Vintedois Diffusion v0.1 | 1.3562 | 1.0797 | 4.6200 | 4.9500 |
| IF-I-L v1.0 | 1.2635 | 0.8814 | 5.2300 | 5.4500 |
| MultiFusion | 1.2372 | 1.1298 | 4.6800 | 4.8000 |
| IF-I-M v1.0 | 1.0135 | 0.7928 | 5.0800 | 5.2200 |