CRAFT-TIDE / README.md
maryanlup's picture
Update README.md
e9e33f3 verified
|
Raw History Blame Contribute Delete
1.83 kB
metadata
license: cc0-1.0
task_categories:
  - text-to-image
language:
  - en
tags:
  - typography
  - font
  - text-rendering
  - benchmark
pretty_name: CRAFT & TIDE
size_categories:
  - 1K<n<10K
configs:
  - config_name: craft
    data_files:
      - split: test
        path: craft.txt
  - config_name: tide
    data_files:
      - split: test
        path: tide.txt

CRAFT & TIDE: Typography Evaluation Benchmarks

Paper · arXiv · Code

Benchmarks for font-conditioned text rendering in text-to-image models, released with FontFusion: Enhancing Generative Text in Diffusion Models with Typographic Conditioning (ICANN 2026).

Config Prompts Avg. words/text Focus
craft 1,605 1.29 Precise font fidelity in minimalist layouts
tide 100 4.19 (137 quoted texts) Realistic multi-word design complexity
from datasets import load_dataset
craft = load_dataset("maryanlup/CRAFT-TIDE", "craft", split="test")
print(craft[0]["text"])

Metrics: OCR word/character accuracy and FontCLIP font consistency. See the GitHub README for the full protocol and results.

Citation

@inproceedings{lupascu2026fontfusion,
    author    = {Lupa\c{s}cu, Marian and Jindal, Nipun and Mironic\u{a}, Ionu\c{t} and Wang, Zhaowen},
    title     = {FontFusion: Enhancing Generative Text in Diffusion Models with Typographic Conditioning},
    booktitle = {Artificial Neural Networks and Machine Learning -- ICANN 2026},
    series    = {Lecture Notes in Computer Science},
    pages     = {40--52},
    publisher = {Springer},
    year      = {2026},
    doi       = {10.1007/978-3-032-38407-2_4}
}