Learnings from training a font recognition model from scratch

35 points by justswim 7 days ago | 7 comments

Terretta 6 days ago |
Instructive, with a rewarding repo for your time.
Tommix11 2 days ago |
Graphic designer here. A font recognition ai is sorley needed. Gemini and its competitors flat out lie when asked and Adobe Illustrators Retype is laughably bad. The problem I face almost every day is not to find a close match but to find the actual font in use, commercial or not.
WillAdams 2 days ago |
whatthefont and identifont used to work well, but they've been overwhelmed by new designs which are not used often enough to warrant inclusion.
I just use Rookledge's Type Finder and a battered copy of Precision Type 5.0
hank_z 2 days ago |
I’m curious to know how this model tackles newly released fonts. How difficult would it be if this model needs to recognise font in a different language?
codemog 2 days ago |
Curious why the model architecture wasn’t talked about at all? Did I miss that part?
g58892881 2 days ago |
pretty basic
``` from torchvision import models
# Avoid downloading pretrained weights; we load trained checkpoint weights. model = models.resnet18(weights=None) model.fc = nn.Linear(model.fc.in_features, num_classes) ```
https://github.com/mixfont/lens/blob/main/lens_inference.py#...
ZeroGravitas 2 days ago |
Surely the next step is to tweak the open source font with AI to slightly better target the desired font?
Sometimes you're trying to replicate a logo and they've hand tweaked a letter or two, or stretched it.
edit: (Google Fonts actually have fonts you can natively stretch and some with replacement character variants) I wonder if the tool uses that to better match?)
Sometimes you just don't want to pay for a fancy font and I believe font shapes are not copyrightable?