Class-conditional image generation on ImageNet 256x256 2015 (test)
2.58FIDNAR
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| NARModel size=XXL (~1.4B), Type=From scratch, Training Epoch=300, Steps=312026.05 | 2.58 | 293.5 | 67.3 | 56.9 | |
| NARModel size=XL (~700M), Type=From scratch, Training Epoch=300, Steps=312026.05 | 2.7 | 277.5 | 67.6 | 98.1 | |
| FlashARModel size=XXL (~1.4B), Type=Post-training, Training Epoch=25, Steps=312026.05 | 2.79 | 289.4 | 69 | 63.4 | |
| FlashARModel size=XL (~700M), Type=Post-training, Training Epoch=25, Steps=312026.05 | 2.94 | 293.7 | 67.2 | 109.3 | |
| NARModel size=L (~360M), Type=From scratch, Training Epoch=300, Steps=312026.05 | 3.06 | 263.9 | 64.1 | 195.4 | |
| LlamaGenModel size=XXL (~1.4B), Type=From scratch, Training Epoch=300, Steps=2562026.05 | 3.09 | 253.6 | 64.7 | 14.1 | |
| FlashARModel size=L (~360M), Type=Post-training, Training Epoch=25, Steps=312026.05 | 3.16 | 289 | 65.6 | 224.7 | |
| PARModel size=XXL (~1.4B), Type=From scratch, Training Epoch=300, Steps=672026.05 | 3.2 | 288.3 | 63.2 | 33.9 | |
| VARModel size=L (~360M), Type=From scratch, Training Epoch=200, Steps=102026.05 | 3.3 | 274.4 | 63.4 | 129.3 | |
| LlamaGenModel size=XL (~700M), Type=From scratch, Training Epoch=300, Steps=2562026.05 | 3.39 | 227.1 | 64.8 | 23.7 | |
| PARModel size=XL (~700M), Type=From scratch, Training Epoch=300, Steps=672026.05 | 3.5 | 234.4 | 61.9 | 53.9 | |
| BlockDiffusionModel size=XXL (~1.4B), Type=Post-training, Training Epoch=75, Steps=642026.05 | 3.78 | 264.9 | 65.2 | 26.8 | |
| LlamaGenModel size=L (~360M), Type=From scratch, Training Epoch=300, Steps=2562026.05 | 3.8 | 248.3 | 63.9 | 47.1 | |
| BlockDiffusionModel size=XL (~700M), Type=Post-training, Training Epoch=75, Steps=642026.05 | 4.13 | 258.6 | 65.4 | 41.7 | |
| PARModel size=L (~360M), Type=From scratch, Training Epoch=300, Steps=672026.05 | 4.32 | 189.4 | 57.6 | 93.8 | |
| BlockDiffusionModel size=L (~360M), Type=Post-training, Training Epoch=75, Steps=642026.05 | 4.55 | 243.5 | 64.5 | 103.2 | |
| NARModel size=B (~120M), Type=From scratch, Training Epoch=300, Steps=312026.05 | 4.65 | 212.3 | 60 | 419.7 | |
| FlashARModel size=B (~120M), Type=Post-training, Training Epoch=25, Steps=312026.05 | 4.68 | 208.3 | 60.5 | 447.2 | |
| LlamaGenModel size=B (~120M), Type=From scratch, Training Epoch=300, Steps=2562026.05 | 5.46 | 193.6 | 59.4 | 117.9 | |
| BlockDiffusionModel size=B (~120M), Type=Post-training, Training Epoch=75, Steps=642026.05 | 5.91 | 176.2 | 58.9 | 186.3 | |
| PARModel size=B (~120M), Type=From scratch, Training Epoch=300, Steps=672026.05 | 6.21 | 204.4 | 53.7 | 174.1 |