search

yeungchenwa / FontDiffuser

[AAAI2024] FontDiffuser: One-Shot Font Generation via Denoising Diffusion with Multi-Scale Content Aggregation and Style Contrastive Learning
https://yeungchenwa.github.io/fontdiffuser-homepage/
230 stars 21 forks source link

About how to generate an SCR_210000.pth file #40

Open RossGeller12138 opened 4 months ago

RossGeller12138 commented 4 months ago

Hello, how was the SCR_210000 used during the second training process saved? How can I separate the corresponding SCR from the total_model. pth file generated during my training? Looking forward to your answer, thank you very much!!!!

yeungchenwa commented 4 months ago

hi@RossGeller12138. SCR is loaded as the style extractor module and is frozen during phase 2 training. Before that, SCR should be pretrained by the fonts (more details are presented in paper appendix).

yeungchenwa commented 4 months ago

hi@RossGeller12138. Feel sorry that the previous codes may confuse you, and I update the code. During phase 2, the trained parameters obtained in phase 1 will be resumed in phase 2. You can turn phase_1_ckpt_dir to your checkpoint files directory obtained in phase 1.

RossGeller12138 commented 4 months ago

Thank you very much for your answer. I will try again according to your prompt. Thank you again for your answer

yeungchenwa commented 4 months ago

I appreciate.

RossGeller12138 commented 4 months ago

Does your current code still not support saving SCR models? Is it only allowed to load and download your SCR model? Can't we train and generate SCR models ourselves? Looking forward to your answer

yeungchenwa commented 4 months ago

Yep. we will release the pre-training script soon~

AlienKevin commented 6 days ago

@yeungchenwa Could you please release the pre-training script? I tested the existing SCR module on handwritten fonts and found that it had a hard time telling apart the different styles. It seems to output one handwritten style consistently, even though differently styled handwritten images were passed in. I'm interested in training a new SCR from scratch with handwritten fonts to see if this can be improved.