Typeorem backbone (DINOv2 ViT-S/14, fp16 ONNX)
The frozen image encoder behind the custom-font feature on
typeorem.dev: DINOv2 ViT-S/14 (timm
vit_small_patch14_dinov2.lvd142m) exported to ONNX, fp16, with ImageNet
normalization baked into the graph.
- Input
pixels: uint8 RGB, shape[batch, 3, 224, 224], values 0..255. - Output
embedding: float32 pooled features, shape[batch, 384].
The browser renders a font's glyph grid to 224x224, runs this graph via
onnxruntime-web, then projects the 384-dim output into the catalog's 200-dim
space (site/data/pca.bin). The graph is produced by model/export_onnx.py --fp16 and published by model/publish_model.py in the parent repository,
https://gitlab.com/fuzzyporpoise/typeorem.
License and attribution
This artifact is a derivative of DINOv2 (Meta AI), distributed under Apache-2.0; the upstream weights and license govern. See https://github.com/facebookresearch/dinov2 and https://huggingface.co/facebook/dinov2-small. The export tooling and the rest of the Typeorem project are MIT.