Image Generation on ImageNet 64x64 (train)
1.23FIDRIN
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| RIN# Fwd Pass=1000, Batch Size=1024, Iterations=300K, Iterated M-images=307.2, Training Hardware=32×TPUv32026.04 | 1.23 | — | — | — | — | |
| EDM (Teacher, SDE)# Fwd Pass=511, Batch Size=4096, Iterations=600K, Iterated M-images=2457.6, Training Hardware=32×A1002026.04 | 1.36 | — | — | — | — | |
| DMD2 (Adv loss)# Fwd Pass=1, Batch Size=280, Iterations=200K, Iterated M-images=56.0, Training Hardware=7×A1002026.04 | 1.51 | — | — | — | — | |
| StyleGAN-XL# Fwd Pass=12026.04 | 1.52 | — | — | — | — | |
| CTM# Fwd Pass=1, Batch Size=2048, Iterations=30K, Iterated M-images=61.4, Training Hardware=8×A1002026.04 | 1.92 | — | — | — | — | |
| ADM#Params=296M, GFLOPS=1102022.09 | 2.07 | 4.29 | — | 74 | 63 | |
| ADM# Fwd Pass=250, Batch Size=768, Iterations=2000K, Iterated M-images=1536.0, Training Hardware=8×V1002026.04 | 2.07 | — | — | — | — | |
| DMD+EL (ours) (Emd loss)# Fwd Pass=1, Batch Size=16, Iterations=200K, Iterated M-images=3.2, Training Hardware=1×RTX40902026.04 | 2.25 | — | — | — | — | |
| EDM (Teacher, ODE)# Fwd Pass=511, Batch Size=4096, Iterations=600K, Iterated M-images=2457.6, Training Hardware=32×A1002026.04 | 2.32 | — | — | — | — | |
| DMD (Reg loss)# Fwd Pass=1, Batch Size=336, Iterations=350K, Iterated M-images=117.6, Training Hardware=7×A1002026.04 | 2.62 | — | — | — | — | |
| iCT-deep# Fwd Pass=1, Batch Size=4096, Iterations=800K, Iterated M-images=3276.8, Training Hardware=N×A1002026.04 | 3.25 | — | — | — | — | |
| BigGAN-deep# Fwd Pass=1, Batch Size=2048, Iterations=200K, Iterated M-images=409.6, Training Hardware=8×TPUv32026.04 | 4.06 | — | — | — | — | |
| U-ViT-L/4Objective=VP, Training iterations=300K, Convolution block=with conv, #Params=287M, GFLOPS=772022.09 | 4.26 | 3.77 | 40.66 | 71 | 62 | |
| Diff-Instruct# Fwd Pass=1, Batch Size=96, Training Hardware=8×V1002026.04 | 5.57 | — | — | — | — | |
| U-ViT-M/4Objective=VP, Training iterations=300K, Convolution block=with conv, #Params=131M, GFLOPS=352022.09 | 5.85 | 4.09 | 33.71 | 69 | 61 | |
| Consistency Model# Fwd Pass=1, Batch Size=2048, Iterations=600K, Iterated M-images=1228.8, Training Hardware=64×A1002026.04 | 6.2 | — | — | — | — | |
| TRACT# Fwd Pass=1, Batch Size=512, Iterations=125K, Iterated M-images=64.0, Training Hardware=8×A1002026.04 | 7.43 | — | — | — | — | |
| Meng et al.# Fwd Pass=1, Batch Size=5122026.04 | 7.54 | — | — | — | — | |
| DFNO# Fwd Pass=1, Batch Size=2048, Iterations=400K, Iterated M-images=819.22026.04 | 7.83 | — | — | — | — | |
| Progress. Distill.# Fwd Pass=1, Batch Size=2048, Iterations=550K, Iterated M-images=1126.4, Training Hardware=8×TPUv42026.04 | 15.39 | — | — | — | — | |
| BOOT# Fwd Pass=1, Batch Size=1024, Iterations=300K, Iterated M-images=307.2, Training Hardware=8×A1002026.04 | 16.3 | — | — | — | — |