The model is trained for 10 epochs with a batch size of 2. Overall, the model is trained for 500,000 iterations, with generation time step T = 200, learning rate 0.0001.
During theinference, generation time steps are set to T = 500. The naturalization of images is donewith the ImageJ Mosaic Suite plugin.
The implementation of CVDM can be found on the GitHub page (G. Della Maggiora, L. A. Croquevielle, N. Deshpande, H. Horsley, T. Heinis, A. Yaki-movich, Conditional variational diffusion models, https://github.com/casus/cvdm)2023.
Further information and documentation (including a guide on installation of the Mosaic Suite plugin) can be found in MOSAIC group, MosaicSuite documentation, https://sbalzarinilab.org/MosaicSuiteDoc/index.html.