Seal generation models
Summary
Motivation
This repository includes all models trained with StyleGAN3 (PyTorch/NVIDIA) on digitized images of imprint seals from the National Archives of France ( See the section below for more details on dataset) collections. These experiments were conducted for educational and research purposes at for applications such as data augmentation, image retrieval, fakes creation for historical studies etc.
Dataset & Model description
Dataset
The training dataset contains 6031 digitized images of imprint seals (accessible from the portal of the National Archives of France) which extend over the period 457-1852. The seals are divided into the following collections:
- Collection Douët d’Arcq Tome 1 : Rois et reines, grands dignitaires, grands feudataires : 1733 images on period 457-1852
- Moulages de sceaux de la collection Lorraine : 1360 images on period 1026-1759
- Moulages de sceaux de la collection du Supplément : 2938 images on period 1300-1700
The images were scraped automatically via a script.
reals seals samples :
fake seals samples during training :
Models
Training was done on NVIDIA GeForce RTX 2080 Ti.
Currently, This repository contains three different models :
| Model name | Config | image size input | GPUs | Batch size | Gamma | Final tick | Final Fid50k_full score |
|---|---|---|---|---|---|---|---|
| seals-stylegan2-256x256_v1.pkl | stylegan2 | 256x256 | 1 | 4 | 6.6 | 920 | 9.46 |
| seals-stylegan2-128x128_v1.pkl | stylegan2 | 128x128 | 1 | 4 | 6.6 | 800 | 7.93 |
| seals-stylegan3-r-256x256_v1.pkl | stylegan3-r | 256x256 | 1 | 4 | 6.6 | 880 | 9.39 |