search

apple / ml-mdm

Train high-quality text-to-image diffusion models in a data & compute efficient manner
https://machinelearning.apple.com/research/matryoshka-diffusion-models
MIT License
438 stars 27 forks source link

[Question] Training another layer #4

Closed Sys-if closed 2 months ago

Sys-if commented 3 months ago

Hi,

First of all, thank you for sharing your code! I've been studying your paper and code, but I’m a bit unclear on how you implemented progressive training. Could you provide some insights into this? For example, do you freeze the inner U-Net for a certain number of iterations?

Sys-if commented 2 months ago

I'm replying to my own comment, as it might be useful for others. I have been experimenting with my own model and have implemented what I understood from the paper by progressively introducing the loss of the new high-resolution layer into the general loss, with a weight increasing from 0 to X. I also used a similar weight to progressively introduce the new layer features (as in Progressive Growing of GANs).