Details of the random transforms for creating user sketch data

[harpavatkeerti]

Hi, really interesting work. I was trying to train the model on a custom dataset, by generating user sketches using the method you mentioned. I couldn't find the implementation details for the set of strong data augmentations (random thresholds, randomly masking out a random percentage of scribbles, random morphological transformations, and random non-maximum suppression) you mentioned in the paper. It would be of great help if you could point me toward that code. Thanks.

[rems75]

@lllyasviel Same question here. I've seen the nms in https://github.com/lllyasviel/ControlNet/blob/main/gradio_fake_scribble2image.py, but it doesn't seem to include the other three augmentations?

[lilisierrayu]

@lllyasviel Could you please kindly share the details.