New in this version:
Datasets
Create and manage datasets for training. You can crop the images and create captions for each of them individually or in bulk. Additionally, the option to use AI (FuseCap) for automatic captioning has been added. However, this should be used as a guide or aid, as the quality of a dataset with auto-generated captions is typically poor.
Training
A new module for training has been added (currently, Dreambooth LoRA). This module allows you to use the created datasets to train a LoRA. All of this can be done without the need to use the command line or connect to an external source.
Bucketed training wasn’t incorporated, as its value wasn’t clear. The models seem to train effectively using the default square ratio. Also, dealing with images of sizes up to 512x2048 presents no issues.
The inclusion of repeat images or data augmentation was not seen as beneficial. Repeating images equates to performing more epochs without augmentation. Augmentation, which often requires bigger size images and bucketing, often results in captions that don’t align if the images are cropped in areas being described by the caption. This can reduce the quality of the dataset. These features may be of interest to those who wish to experiment. However, for effective fine-tuning or LoRA, manual intervention is necessary, which aligns with the purpose of this software.
Exciting enhancements are on the horizon!
- IP Adapters: Add even more tools to your arsenal with this powerful adapters.
- Image to Image: Transform your visual content with our soon-to-be-launched Image to Image feature.
- Inpainting: Experience the magic of seamless photo editing with our forthcoming Inpainting tool.
Stay tuned for these game-changing additions that will take your productivity to new heights!