search

yizhiwang96 / deepvecfont

[SIGGRAPH Asia 2021] DeepVecFont: Synthesizing High-quality Vector Fonts via Dual-modality Learning
MIT License
182 stars 31 forks source link

Random Sampling From Latent Space #14

Open ZloyTapok987 opened 2 years ago

ZloyTapok987 commented 2 years ago

Hi! Can you share code for Random Sampling From Latent Space experiment?

ZloyTapok987 commented 2 years ago

I'm trying to sample new font via random latent space with this code: https://pastebin.com/m8vyYpXx But it doesn't work, svg images don't looks like glyphs :( What I'm doing wrong

Screenshot 2022-04-27 at 01 48 50

?

yizhiwang96 commented 2 years ago

Hi Zloy！ Thanks for your questions! Yes, I re-ran my codes and also met your problem. However, I found font-interpolation works well. Then, I checked my TensorFlow codes (which were used to produce the results in my paper.) I find kl_beta was set to 0.1 in my TF codes. I think this could be the reason. I'm busy with my thesis recently and I will try to re-run my Pytorch code with kl_beta=0.1 (or find some other reasons). It would be appreciated if you help to check if set kl_beta=0.1 works (maybe some other weights need to be finetuned accordingly). Thank you!

cravisjan97 commented 2 years ago

Hi! I am also facing the same issue. Did changing the KL beta weight work for random font generation?

yizhiwang96 commented 2 years ago

kl_beta=0.1

Hi Cravis, Yes, kl_beta=0.1 works for random font generation. I also noticed that kl_beta=0.01 results in better reconstruction than kl_beta=0.1. So I think it is a kind of tradeoff and I will upload a checkpoint of kl_beta=0.1.