Request for English IAM-OnDB pretrained weights

[gcharycki]

Hi,

Thank you for your work on DiffInk.

I am currently working on a multilingual extension of the model for English and Polish (using ~11k IAM-OnDB samples + ~3.5k Polish data).

To debug my pipeline, I attempted to reproduce your English baseline results (Appendix D.3). I kept all hyperparameters consistent, only adjusting the base_lr to 5.0e-5 for InkDiT training as specified. Despite this, my results are significantly below your benchmarks.

Could you please share your English (IAM-OnDB) pretrained weights? Having your baseline checkpoint would be invaluable to verify if my issues stem from the training configuration or the multilingual setup.

Thank you for your time and your research!

Best regards,
gcharycki

[awei669]

Dear gcharycki,

Thank you for your interest in our DiffInk method and for your efforts in reproducing the English benchmark results.

The current repository is mainly optimized for the Chinese handwriting generation setting, and directly applying it to English or multilingual training may lead to some issues. We have recently been integrating and improving the codebase for multilingual scenarios, including English or other languages.

The corresponding code and pretrained weights will be made publicly available once the training and validation are completed. If there is any progress, I will let you know as soon as possible.

Thank you again for your attention to our work.

Best regards,
Wei

[gcharycki]

Dear Wei,

Thank you for the clarification. I look forward to the updated codebase and weights.

Best regards,
gcharycki