Image Generation on ImageNet 512x512 (train)
3.85FIDADM-G, ADM-U
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| ADM-G, ADM-U#Params=422M + 43M (Cls) + 309M (SR), GFLOPS=307 + 21 (Cls) + 2506 (SR)2022.09 | 3.85 | 5.86 | 221.72 | 84 | 53 | |
| U-ViT-H/4Objective=SD, Training iterations=500K, Convolution block=without conv, Weight mode=ft-EMA, p_uncond=0.1, #Params=501M + 84M (AE), GFLOPS=133 + 1260 (AE)2022.09 | 4.05 | 6.44 | 263.79 | 84 | 48 | |
| U-ViT-H/4Objective=SD, Training iterations=500K, Convolution block=without conv, Weight mode=original, p_uncond=0.1, #Params=501M + 84M (AE), GFLOPS=133 + 1260 (AE)2022.09 | 4.34 | 8.44 | 261.13 | 84 | 48 | |
| U-ViT-L/4Objective=VP, Training iterations=500K, Convolution block=with conv, Weight mode=original, p_uncond=0.15, #Params=287M + 84M (AE), GFLOPS=77 + 1260 (AE)2022.09 | 4.67 | 5.87 | 213.28 | 87 | 45 | |
| U-DiT-BFLOPS (G)=106.7, Training Iterations=400K2024.05 | 15.39 | 6.86 | 92.73 | 75.6 | 60.5 | |
| DiT-XL/2*FLOPS (G)=524.7, Training Iterations=400K2024.05 | 20.94 | 6.78 | 66.3 | 74.5 | 58.1 |