I give a summary about the training of the text-to-video model: they used 48 GPUs for the keyframes model 128k steps, 24 GPUs for the interpolation model 14k st

I give a summary about the training of the text-to-video model: they used 48 GPUs for the keyframes model 128k steps, 24 GPUs for the interpolation model 14k steps, 32 GPUs for the upscaler model 10k steps, 40 GPUs for the autoencoder 35k steps. They used A100-80G in all cases except the autoencoder (A100-40G). This is not particularly expensive considering the numbers these companies usually use, thanks to the fact that they only trained the temporal layers in the LDM and the autoencoder.