why cond_stage_trainable is set true?

XavierXiao / Dreambooth-Stable-Diffusion

Implementation of Dreambooth (https://arxiv.org/abs/2208.12242) with Stable Diffusion

MIT License

7.6k stars 795 forks source link

why cond_stage_trainable is set true? #74

Open liuyuyuil opened 2 years ago

liuyuyuil commented 2 years ago

I have found that the hyper-parameter cond_stage_trainable is set true in the config file. The cond_stage_model should be the CLIP text encoder, why fintuning the parameters of this module is necessary? In the original paper, have they finetuned the T5 encoder? Thanks!

jshi31 commented 1 year ago

I also have the same question. Seems that it finetunes all the UNet parameters and the CLIP text encoder parameters.