Conversational Image Editing on OCCUR-Bench 4,400 scenarios (test)
61.6Restoration ScoreGemini-2.5 w/ ReSpec
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| Gemini-2.5 w/ ReSpecHistorical Reference Selection=false, sample size=200-sample subset2026.07 | 61.6 | 58 | 59.8 | 76.6 | |
| Flux.2 w/ ReSpecReSpec framework=true2026.07 | 54.7 | 62 | 58.4 | 75.6 | |
| Gemini-2.5sample size=200-sample subset2026.07 | 53.8 | 54.4 | 54.1 | 60.5 | |
| Flux.22026.07 | 38.6 | 52.5 | 45.5 | 65.2 | |
| OmniGen2 w/ ReSpecReSpec framework=true2026.07 | 38.1 | 42.3 | 40.2 | 69.1 | |
| MC-Edit2026.07 | 32.3 | 36.7 | 34.5 | 27.1 | |
| OmniGen22026.07 | 32 | 42 | 37 | 56.1 | |
| Layer-wise Memory2026.07 | 29.1 | 29 | 29 | 27.3 |