Image Generation on ImageNet 256x256 (400K iterations test)
11.36FIDDiC-H
Evaluation results
| Method | Links | ||
|---|---|---|---|
| DiC-HFLOPs (G)=204.4, Winograd FLOPs (G)=97.2, TP=160.8, Training Steps=400K2024.12 | 11.36 | 106.52 | |
| DiC-XLFLOPs (G)=116.1, Winograd FLOPs (G)=57.2, TP=313.7, Training Steps=400K2024.12 | 13.11 | 100.15 | |
| U-ViT-XLFLOPs (G)=113, TP=72.6, Training Steps=400K2024.12 | 18.35 | 76.59 | |
| DiT-XL/2FLOPs (G)=118.6, TP=66.8, Training Steps=400K2024.12 | 20.05 | 66.74 | |
| DiT-LLaMA*FLOPs (G)=118.6, TP=65.2, Training Steps=400K2024.12 | 20.22 | 70.1 | |
| PixArt-α-XL/2FLOPs (G)=118.4, TP=64.1, Training Steps=400K2024.12 | 24.75 | 52.24 | |
| DiffiT-XL/2FLOPs (G)=118.5, TP=64.1, Training Steps=400K2024.12 | 36.86 | 35.39 |