TI was trained on a dataset of 32 publicly-available images with a batch size of 12 and gradient accumulation of 4 (16 images were repeated for a total training dataset count of 48) for 130 epochs using the ADAMW scheduler and a constant LR of 0.003. Image masking was leveraged to focus the training. Tattoos were removed from the training data; scars were not (models may occasionally interpret them as tattoos).
TI was trained on a dataset of 32 publicly-available images with a batch size of 12 and gradient accumulation of 4 (16 images were repeated for a total training dataset count of 48) for 130 epochs using the ADAMW scheduler and a constant LR of 0.003. Image masking was leveraged to focus the training. Tattoos were removed from the training data; scars were not (models may occasionally interpret them as tattoos).