Total training time about diffae HOT 4 CLOSED

This information is in the supplementary material.
You can see the throughput (imgs/gpu/sec) here:

from diffae.

That seems right. However, larger datasets don't necessarily take more time to train, at least not likely in proportion to the size of the dataset. You need a certain number of images, aka. iterations (not unique images) to get a certain quality of diffusion model. It's unlikely that double the dataset size would take 2x more time to train. Besides, you may decide to end the training whenever the quality is acceptable. You may set a fixed number of images like 100M regardless of the number of image in the dataset.

from diffae.

Thanks a lot! So if my understanding and calculation are close to correct, then training could take a little under a week for FFHQ-128 with 4 V100, and with a larger dataset, it could span several weeks. Am I correct?

from diffae.

Thanks, I see now! This is not an easy task to train such models in terms of time and computing, so hoping that in future some possible research in the area of efficiency will pop up :)

from diffae.