Multi-subject Text-to-Image Generation on Orthogonal Adaptation concept bank
75.9Text Alignment Score (Single)LILAC (LaDe)
Evaluation Results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| LILAC (LaDe)Merge time=none, Backbone=LaDe, Configuration=native layered2026.07 | 75.9 | 75.8 | 0.001 | 76.1 | 74.5 | 0.016 | 87.9 | 87.7 | 0.002 | |
| LILAC (decomposition)Merge time=none, Backbone=Qwen-Image-Edit, Configuration=decomposition2026.07 | 75.4 | 73.4 | 0.02 | 74.6 | 74.3 | 0.003 | 96.1 | 63.9 | 0.322 | |
| LILAC (scaffold)Merge time=none, Backbone=Qwen-Image-Edit, Configuration=scaffold2026.07 | 75.4 | 71.1 | 0.043 | 74.6 | 73.8 | 0.008 | 96.1 | 86.1 | 0.1 | |
| Orthogonal Adaptation (Qwen)Merge time=<1 s, Backbone=Qwen-Image-Edit2026.07 | 75 | 66.4 | 0.086 | 75 | 86.9 | 0.119 | 92.9 | 0 | 0.929 | |
| Custom DiffusionMerge time=~2 s, Backbone=SDXL2026.07 | 66.8 | 67.3 | 0.005 | 64.8 | 62.3 | 0.025 | 50.4 | 40.8 | 0.096 | |
| P+Merge time=<1 s, Backbone=SDXL2026.07 | 64.3 | 64.3 | — | 68.3 | 68.3 | — | 51.5 | 51.5 | — | |
| Mix-of-Show (FedAvg)Merge time=<1 s, Backbone=SDXL2026.07 | 62.5 | 62.1 | 0.004 | 74.5 | 73.5 | 0.01 | 72.8 | 70.6 | 0.022 | |
| Mix-of-Show (Grad. Fusion)Merge time=∼15 m, Backbone=SDXL2026.07 | 62.5 | 63.1 | 0.006 | 74.5 | 72.9 | 0.016 | 72.8 | 71.7 | 0.011 | |
| Orthogonal Adaptation (SDXL)Merge time=<1 s, Backbone=SDXL2026.07 | 62.4 | 64.4 | 0.01 | 74.8 | 74.1 | 0.007 | 74 | 74.5 | 0.005 | |
| DB-LoRA (FedAvg)Merge time=<1 s, Backbone=SDXL2026.07 | 61.3 | 68.2 | 0.069 | 74.4 | 53.1 | 0.213 | 68.3 | 9.8 | 0.585 |