Text-to-Image Generation on 40 hand-designed conflicting concept sets (test)
31.13CLIP Joint ScoreDivide-and-Denoise
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| Divide-and-DenoiseCoordination Strategy=Divide-and-Denoise, Base Model=Stable Diffusion2026.06 | 31.13 | 21.23 | 1.12 | -0.55 | 90.5 | 81.5 | |
| MultiDiffusionCoordination Strategy=MultiDiffusion, Base Model=Stable Diffusion2026.06 | 29.85 | 20.23 | 0.7 | -0.85 | 81.9 | 68.9 | |
| Multi-Concept Diff.Coordination Strategy=Multi-Concept Diffusion, Base Model=Stable Diffusion2026.06 | 29.76 | 19.82 | 0.65 | -0.88 | 75.2 | 63.3 | |
| Composable Diff.Coordination Strategy=Composable Diffusion, Base Model=Stable Diffusion2026.06 | 29 | 20.2 | 0.08 | -1.23 | 72.9 | 58.1 | |
| AveragingCoordination Strategy=Averaging, Base Model=Stable Diffusion2026.06 | 27.81 | 19.14 | -0.36 | -1.45 | 68.7 | 52.1 |