In v-prediction, the loss is higher in the early timesteps (near the noise). This option can be used to increase the loss in the early timesteps. tools/cache_latents.py is added. This script can be ...
Some results have been hidden because they may be inaccessible to you
Show inaccessible results
Feedback