search

SHI-Labs / Prompt-Free-Diffusion

Prompt-Free Diffusion: Taking "Text" out of Text-to-Image Diffusion Models, arxiv 2023 / CVPR 2024
https://arxiv.org/abs/2305.16223
MIT License
735 stars 36 forks source link

how to train seecoder? #11

Closed pokameng closed 1 year ago

pokameng commented 1 year ago

@xingqian2018

pokameng commented 1 year ago

hello @xingqian2018 I want to know how to train seecoder

xingqian2018 commented 1 year ago

The details about training the model are in Section 3.3 last paragraph, and Section 4.2 for more detailed hyperparameters. In conclusion, it is like Image-Variation training using the common diffusion loss function (variational-low-bounded negative log-likelihood).

We still do not have any plan to release the training code. Need supervisor's approval.