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/
231 stars 21 forks source link

数据集相关问题 #45

Open Goldatch opened 3 months ago

Goldatch commented 3 months ago

您好,这个项目真的太牛了,我想要自己训练一下,但是不知道这个项目数据集的那424种字体的选取有什么特别的考虑吗。还是说从公开字库里随机下载的这些字体并没有什么考虑。如果是有特别进行挑选的话,是个什么挑选策略呢

yeungchenwa commented 3 months ago

@Goldatch 您好,感谢您对本项目的关注,我们文章提到的数据集对于字体的选择是没有特殊的考虑的,我们在实验过程中是直接在公开字库中进行随机选取的;当然依照我们的经验,当训练数据的字体风格越多样,字的复杂度多样,模型的鲁棒性会更好~

Note-Liu commented 2 months ago

@Goldatch 您好,感谢您对本项目的关注,我们文章提到的数据集对于字体的选择是没有特殊的考虑的,我们在实验过程中是直接在公开字库中进行随机选取的;当然依照我们的经验,当训练数据的字体风格越多样,字的复杂度多样,模型的鲁棒性会更好~

When using TTF to generate font images, does the absence of certain Chinese characters in a TTF file affect training? Must font images for each style be complete? Thx.

yeungchenwa commented 2 months ago

@Goldatch 您好,感谢您对本项目的关注,我们文章提到的数据集对于字体的选择是没有特殊的考虑的,我们在实验过程中是直接在公开字库中进行随机选取的;当然依照我们的经验,当训练数据的字体风格越多样,字的复杂度多样,模型的鲁棒性会更好~

When using TTF to generate font images, does the absence of certain Chinese characters in a TTF file affect training? Must font images for each style be complete? Thx.

Hi @Note-Liu It is no need to enable the font images for each style be complete.