Repo for controlnet model trained on 3.5M samples from coyo-700M dataset and 2.6M samples from bridge dataset
For each training step, the batch is composed of 28 images from coyo and 4 from Bridge.
We created a custom DataLoader that would load the batch with this 28:4 ratio each step from randomly selected images in both dataset.
Therefore, we didn't trained for a certain number of epoch but for a certain number of step.
If we consider that an epoch correspond to the process of 2.6M images, then we processed 2.275M images from coyo and 325k images from bridge per epoch.