GenImage ADM accuracy much lower than reported result

[makabakallllll]

Hi, thanks for open-sourcing this great work! I'm excited to try it out.

I ran the code on the GenImage dataset using a single RTX 4090, following the same configuration as the provided train.sh — training on stable_diffusion_v_1_4 and evaluating on all other subsets.

The paper reports ADM accuracy at 96.68% on GenImage, but I'm only getting ~72%.

Could you help clarify:

Are there any differences between the released code and the training config in train.sh that could explain this gap?

Does the paper's GenImage result use a different embedding precomputing recipe?

train epoch 5:
                subset  accuracy  real_accuracy  fake_accuracy       ap       f1
                   ADM  0.709417       0.954333       0.464500 0.849598 0.615164
                BigGAN  0.879917       0.959167       0.800667 0.962611 0.869581
                 glide  0.927583       0.953333       0.901833 0.982136 0.925669
            Midjourney  0.931667       0.952833       0.910500 0.984518 0.930189
stable_diffusion_v_1_4  0.976417       0.959500       0.993333 0.998553 0.976809
                  VQDM  0.931417       0.956667       0.906167 0.982783 0.929640
                wukong  0.970417       0.958333       0.982500 0.996912 0.970770
             Chameleon  0.832367       0.745744       0.947628 0.961499 0.829091

train epoch 10:
                subset  accuracy  real_accuracy  fake_accuracy       ap       f1
                   ADM  0.721833       0.948167       0.495500 0.852250 0.640457
                BigGAN  0.896167       0.950167       0.842167 0.968281 0.890240
                 glide  0.934583       0.944667       0.924500 0.984731 0.933917
            Midjourney  0.934917       0.947667       0.922167 0.984949 0.934076
stable_diffusion_v_1_4  0.970000       0.945000       0.995000 0.998506 0.970732
                  VQDM  0.937750       0.948833       0.926667 0.985311 0.937052
                wukong  0.968083       0.949667       0.986500 0.997146 0.968660
             Chameleon  0.818269       0.714190       0.956759 0.962727 0.818770