search

tlc121 / FsFont

Official PaddlePaddle Implementation of Few-Shot Font Generation by Learning Fine-Grained Local Styles (FsFont)
Other
64 stars 8 forks source link

Some questions about Ablation studies in the paper. #1

Open mcshih opened 1 year ago

mcshih commented 1 year ago

Hi! Thanks for sharing the great work! I have some questions about FsFont. Compare Ablation part and Experimental results, the model without any new module (the last row in Ablation studies table, which can I say it is just a GAN with reference encoder & content encoder ) is still has quite good performance even compared to LF-Font or MX-Font. Did I miss any details? fsfont_1 fsfont_0

tlc121 commented 1 year ago

Hi! Thanks for sharing the great work! I have some questions about FsFont. Compare Ablation part and Experimental results, the model without any new module (the last row in Ablation studies table, which can I say it is just a GAN with reference encoder & content encoder ) is still has quite good performance even compared to LF-Font or MX-Font. Did I miss any details? fsfont_1 fsfont_0

Thanks for your attention! You are right, the last row is just a GAN with two encoders and a decoder but we fixed the number of references in 3 for each content character without Content-Reference mapping. Both of LFFont and MXFont is trying to explicitly disentangle the content and style information in networks, so they have a quiet good performence in the Character Accuracy which means they can generate a very stable font but both of them will lose many style information. We've already clarified this point in the Abstract and Section 2: "Style lies in the local details". Those 4 Evaluation metrics in ablation studies is to see how similar between the output and ground truth. Actually the last row setting will have some defects like a dot in blank or missing a stroke but these defects will not have many impacts on these Evaluation metrics. You can do an experiment on Character Accuracy in the same experiment setting.