Text-to-image Generation on LAION Aesthetics 6.5+ 512x512 zero-shot on MS-COCO V2
13.93FIDTeacher
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| TeacherGen. Type=Diffusion, Architecture=SD 2.1, Iters=1620k, MACs=339.2G, Params=865.9M2026.05 | 13.93 | 35.93 | 0.3075 | |
| LIFT + PLACEGen. Type=Diffusion, Architecture=BK-Tiny-v2, Iters=50k, MACs=205.3G, Params=326.8M, K=162026.05 | 14.6 | 31.75 | 0.2916 | |
| LIFT + PLACEGen. Type=Diffusion, Architecture=BK-Small-v2, Iters=50k, MACs=218.0G, Params=485.8M, K=162026.05 | 14.65 | 31.99 | 0.2918 | |
| BK-SDMGen. Type=Diffusion, Architecture=BK-Tiny-v2, Iters=50k, MACs=205.3G, Params=326.8M2026.05 | 15.68 | 31.64 | 0.2897 | |
| BK-SDMGen. Type=Diffusion, Architecture=BK-Base-v2, Iters=50k, MACs=224.1G, Params=583.5M2026.05 | 15.85 | 31.7 | 0.2868 | |
| BK-SDMGen. Type=Diffusion, Architecture=BK-Small-v2, Iters=50k, MACs=218.0G, Params=485.8M2026.05 | 16.61 | 31.73 | 0.2901 | |
| LIFT + PLACEGen. Type=Diffusion, Architecture=BK-Base-v2, Iters=50k, MACs=224.1G, Params=583.5M, K=162026.05 | 16.72 | 31.82 | 0.29 | |
| TeacherGen. Type=Flow, Architecture=SD 3 (D24), MACs=799.24G, Params=2.03B2026.05 | 20.26 | 40.12 | 0.3089 | |
| LIFT + PLACEGen. Type=Flow, Architecture=SD 3 (D21), Iters=50k, MACs=699.08G, Params=1.77B, K=16, Initialization=ShortGPT [23]2026.05 | 20.59 | 37.03 | 0.3022 | |
| OutKD+FeatKDGen. Type=Flow, Architecture=SD 3 (D21), Iters=50k, MACs=699.08G, Params=1.77B, Initialization=ShortGPT [23]2026.05 | 21.17 | 36.82 | 0.303 | |
| LIFT + PLACEGen. Type=Flow, Architecture=SD 3 (D18), Iters=50k, MACs=598.83G, Params=1.52B, K=16, Initialization=ShortGPT [23]2026.05 | 21.21 | 25.14 | 0.2696 | |
| OutKD+FeatKDGen. Type=Flow, Architecture=SD 3 (D18), Iters=50k, MACs=598.83G, Params=1.52B, Initialization=ShortGPT [23]2026.05 | 22.72 | 24.28 | 0.2662 |