Efficient Fine-tuning on SANA
3.1Training Memory (GiB)DiT-BlockSkip
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| DiT-BlockSkipSkip Ratio=50 %, Training Resolution=512×5122026.03 | 3.1 | 1.79 | 0.51 | 0.79 | 0.9 | 0.82 | |
| DiT-BlockSkipSkip Ratio=40 %, Training Resolution=512×5122026.03 | 3.69 | 2.14 | 0.62 | 0.94 | 1.08 | 0.82 | |
| HollowedNetSkip Ratio=50 %, Training Resolution=1024×10242026.03 | 4.26 | 1.79 | 0.51 | 1.95 | 3.19 | 2.91 | |
| DiT-BlockSkipSkip Ratio=30 %, Training Resolution=512×5122026.03 | 4.29 | 2.48 | 0.72 | 1.09 | 1.26 | 0.82 | |
| HollowedNetSkip Ratio=40 %, Training Resolution=1024×10242026.03 | 5.08 | 2.14 | 0.62 | 2.33 | 3.84 | 2.91 | |
| HollowedNetSkip Ratio=30 %, Training Resolution=1024×10242026.03 | 5.9 | 2.48 | 0.72 | 2.7 | 4.48 | 2.91 | |
| LISATraining Resolution=1024×10242026.03 | 6.67 | 3.5 | 0.1 | 3.07 | 6 | — | |
| LoRA-FATraining Resolution=1024×10242026.03 | 7.4 | 3.5 | 0.51 | 3.39 | 6.16 | — | |
| LoRATraining Resolution=1024×10242026.03 | 8.35 | 3.5 | 1.03 | 3.83 | 6.41 | — | |
| Textual InversionTraining Resolution=1024×10242026.03 | 15.31 | 7.86 | 2.2 | 5.25 | 8.91 | — | |
| DreamBoothTraining Resolution=1024×10242026.03 | 24.69 | 2.99 | 5.98 | 15.72 | 13.34 | — |