Training Set Consultation

[jingmingtao]

Hello: The training set consists of content and style, with 6763 characters for content and 556 characters for style. After training and testing, it was found that the character style generation effect was not good except for 556. May I ask if this dataset combination is reasonable

[jingmingtao]

I want to use a small number of style fonts to generate a reference style font library, so the training set created in this way

[8uos]

Note that, our method is not a few-shot training method. Our method assumes that the model can access large number of samples during training. How many style fonts (not characters) are you using? If you are trying to train our model only with one style font, stop it and check the inference.ipynb.

[jingmingtao]

Thank you for your reply！！！ My requirement: Suppose I have a style font A with only 556 characters (designed by the designer), and I want to generate the remaining 6763 characters of font A. How should I do it? The method I can think of is to use the known 556 characters as the style, and then find a 6763 character font B as the content for training. This will generate 6763 for font A. But your answer made me realize that my approach seems incorrect. So, if small sample training is not possible, how can untrained small samples generate a large number of font libraries and generate the remaining 6763 characters of font A?

------------------ 原始邮件 ------------------ 发件人: "clovaai/mxfont" @.>; 发送时间: 2023年11月21日(星期二) 中午11:49 @.>; @.@.>; 主题: Re: [clovaai/mxfont] Training Set Consultation (Issue #29)

Note that, our method is not a few-shot training method. Our method assumes that the model can access large number of samples during training. How many style fonts (not characters) are you using? If you are trying to train our model only with one style font, stop it and check the inference.ipynb.

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>