4.5.2. Individual Evaluation

LTTR24 Base Model

The model’s variant, trained on 500 epochs, provided the best results on all three efforts at aesthetics heuristics. The version on 300 epochs performed slightly better results than on 600 epochs on detail precision and alphabet consistency.

The number of generated samples has not produced significantly different results across all the epoch variants.

Figure screenshot LTTR24 Base Model. Red layer: 300 epochs, Blue layer: 500 epochs

DeepVecFont-2 Original Model

In the experiment, only one model’s version was tested. Therefore, training effort results were not evaluated.

Figure screenshot DeepVecFont-2 Original Model. Red layer: 50 samples, Blue layer: 100 samples

From the inference effort perspective, the Reference matching heuristic does seem to provide some difference, where 50 and 100 generated samples provided better results than 20 and 30 generated samples. However, this difference could be questionable.

Figure screenshot DeepVecFont-2 Original Model comparing detail reference matching. Black layer: Reference, Aqua layer: 20 samples, Red layer: 30 samples, Lime layer: 50 samples, Blue layer: 100 samples

DeepVecFont-2 Fine-tuned on the LTTRSET Dataset

In the detail precision criterion, the model’s variant trained on 800 epochs performed slightly better results than the 1000 epochs variant. However, the difference is not significant and can be prone to subjective judgement bias.

Figure screenshot DeepVecFont-2 Fine-tuned on the LTTRSET Dataset. Red layer: 800 epochs, Blue layer: 1000 epochs

The reference matching performance seems to be for both variants seem to be similar. In the alphabet consistency criterion, the 1000 epochs variant achieved slightly better results.

Figure screenshot DeepVecFont-2 Fine-tuned on the LTTRSET Dataset comparing reference matching. Black layer: Reference, Red layer: 800 epochs, Blue layer: 1000 epochs

LTTR24 Fine-tuned on the SVG-Font Dataset

The model’s variant, trained on 1100 epochs, provided the best results on all three efforts at aesthetics heuristics. The version on the 1200 epochs performed slightly worse results. The 800 and 1000 epochs variant generated strange images that don’t even look like letters.

The number of generated samples has not produced significantly different results across all the epoch variants.

Figure screenshot LTTR24 Fine-tuned on the SVG-Font Dataset. Lime layer: 800 and 1000 epochs, Red layer: 1100 epochs, Blue layer: 1200 epochs

License

This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License .

Creative Commons License

Citation

If this work is useful for your research, please cite it as:

@phdthesis{paldia2025generative,
  title={Research and development of generative neural networks for type design},
  author={Paldia, Filip},
  year={2025},
  school={Academy of Fine Arts and Design in Bratislava},
  address={Bratislava, Slovakia},
  type={Doctoral thesis},
  url={https://lttrface.com/doctoral-thesis/},
  note={Department of Visual Communication, Studio Typo}
}