search

rekil156 / blogComments

0 stars 0 forks source link

rekilblog/posts/lesson9_stableDissufion/Lesson9 #1

Open utterances-bot opened 1 year ago

utterances-bot commented 1 year ago

rekilblog - A different way to look at Stable Diffusion

https://rekil156.github.io/rekilblog/posts/lesson9_stableDissufion/Lesson9.html

connieK8888 commented 1 year ago

Can the magic function solution generate a chinese calligraphy albums under or presented the specific personalized calligraphy style?

mayankkamboj47 commented 9 months ago

"The encoder of the VAE is only required during training and not during inference."

Did you mean decoder ? The encoder should still be needed because the noise prediction (which we'll try to minimise) works with latents, not images right ? The decoder on the other hand, only is needed to train the noise predictor, not to use it later.

mayankkamboj47 commented 9 months ago

Another question, why not use varying activations in the same neuron rather than a space-inefficient one-hot-encoded vector ? Eg. Activation of 1, 2, 3, 4, etc. each meaning different things rather than using a one-hot-encoded vector. It also saves compute required to train the text encoder