Image Generation Efficiency on ImageNet 256x256 (train)
166.67Training TimeDense-DiT-L-Flow
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Dense-DiT-L-Flow#Experts=-, Activated Params.=458M, Total Params.=458M, Training Steps=500K, cfg=1.52025.10 | 166.67 | 1.25 | 77.5 | |
| EC-DiT-L-Flow#Experts=E8A1S0U0, Activated Params.=458M, Total Params.=1.163B, Training Steps=500K, cfg=1.52025.10 | 277.78 | 1.49 | 77.5 | |
| DiT-MoE-L-Flow#Experts=E8A1S0U0, Activated Params.=458M, Total Params.=1.163B, Training Steps=500K, cfg=1.52025.10 | 333.33 | 1.49 | 77.5 | |
| DiffMoE-L-Flow#Experts=E8A1S0U0, Activated Params.=458M, Total Params.=1.095B, Training Steps=500K, cfg=1.52025.10 | 333.33 | 1.64 | 82.53 | |
| ProMoE-L-Flow#Experts=E14A1S1U1, Activated Params.=458M, Total Params.=1.063B, Training Steps=500K, cfg=1.52025.10 | 333.33 | 1.53 | 77.72 |