EDM: 250/256 steps

[Felix-Petersen]

Hi,

thanks for the great paper and code. I found a potential typo and confusion on a result for EDM with 250/256 samples.

My confusion comes from 250 vs. 256 samples: In C.3 and 5.2, you state 250 samples. But then in C.4 and Table C.1, you say it's 256 samples. Also, the EDM source only provides guidance on 256 samples. Probably a typo and you meant actually always write 256 samples?

Also, rather than including 250/256 samples for EDM, you make a new table in the appendix. Why not include it in the actual paper? I mean the result is quite strong.

Thanks and best regards,

Felix

[Felix-Petersen]

I've attempted to reproduce 256 steps for EDM with --max_guidance 0.001 --steps 256. However, setting --use_classifier 0 leads to substantially better results for me.

For the other numbers of steps, I have observed results matching those reported in the paper.

Are there any other hyperparameters that need to be changed for evaluating with 256 steps?

EDIT: Adding --S_churn 40 --S_min 0.05 --S_max 50 --S_noise 1.003 improves the preliminary results for me for now, but it will take some time until I have the results for the full experiment to see whether this change is enough to get it to work as reported.

[Felix-Petersen]

UPDATE: using --max_guidance 0.001 --steps 256 --S_churn 40 --S_min 0.05 --S_max 50 --S_noise 1.003, I achieved an FID of 1.358 (50k images). Which parameters do I need to get it down to 1.33?

[AlexMaOLS]

Hi,

thanks for the great paper and code. I found a potential typo and confusion on a result for EDM with 250/256 samples.

My confusion comes from 250 vs. 256 samples: In C.3 and 5.2, you state 250 samples. But then in C.4 and Table C.1, you say it's 256 samples. Also, the EDM source only provides guidance on 256 samples. Probably a typo and you meant actually always write 256 samples?

Also, rather than including 250/256 samples for EDM, you make a new table in the appendix. Why not include it in the actual paper? I mean the result is quite strong.

Thanks and best regards,

Felix

Thank you for pointing out the typo! I You are right that I meant 256 steps for EDM SDE sampling. In terms of results SDE results replication, I am not sure if it is related to any other issues. It is convenient for you to check if the guided sampling result for EDM in fewer steps of ODE sampling first? Just to make sure the setting is right. Thank you for your response!

Jiajun