SelfEval: Leveraging the discriminative nature of generative models for evaluation

Benchmark Model Rank Results
visual-reasoning-on-winogroundCLIP (ViT-L/14)–Text Score: 30.25Image Score: 8.0
visual-reasoning-on-winogroundLDM-CLIP (SelfEval)–Text Score: 22.75Image Score: 7.25
visual-reasoning-on-winogroundLDM-T5 (SelfEval)–Text Score: 29.00Image Score: 13.50
visual-reasoning-on-winogroundOCLIP (ViT-H/14)–Text Score: 30.75Image Score: 12.75
visual-reasoning-on-winogroundPDM-CLIP (SelfEval)–Text Score: 17.00Image Score: 14.00
visual-reasoning-on-winogroundPDM-T5 (SelfEval)–Text Score: 28.25Image Score: 12.00