Thank you for your work on this interesting paper. I would like to highlight another issue that we were hoping you could help resolve.
Table 7 of the Appendix states that the number of parameters for the ADM model for CelebA is 368M parameters.
The reported hyperparameters for ADM (CelebA-HQ-256) in Table-7 in the Appendix seem to be different than the hyperparameters in the text file for the corresponding downloadable checkpoint (celeba_f8_adm in the repository). In the Table, the "Channel multipliers per scale" are listed as (1,2,3,4) and the "Attention resolutions" are listed as (16,8,4) , whereas on the text file for the "celeba_f8_adm" checkpoint :
MODEL_TYPE=adm
EPOCH_ID=450
DATASET=celeba_256
EXP=celeb256_f8_adm
METHOD=dopri5
STEPS=0
USE_ORIGIN_ADM=true
IMG_SIZE=256
CH_MULT="1 2 2 2"
ATTN_RES="16 8"
Loading the above checkpoint leads us to count the number of parameters for the ADM model for CelebA to be ~153M parameters.
We would appreciate it if you could confirm the typo in the paper.