- new models for tagging and/or captioning
- new augmentation options
- new image board support (anime & gelbooru)
- two new image tagging modes for both tag based models and caption based models
- a new UI for both modes
- a new layout
- a better way of outlining tags and words in different categories
- faster response times with tag/word suggestions
------------
- segmentation & object detection model/s integration
- a new feature to automatically curate data for the user according to a few questions the user has to answer to determine the heuristics of the model
- cogVLM support
- LLM support for optionally improving captions i.e. cogVLM models struggle with NSFW tagging, so it might be better to use tags resulting from a tagging model and have an LLM re-format the tags into a caption; and then as a second pass have the VLM act as validation to those captions. And/or use an LLM to also merge the LLM caption and the cogVLM caption together into 1 singular caption that takes certain semantics of both captions to make a better one.
- more user configurable options to caption their data
- new methods of pruning tags, again based on various heuristics from Q&A with the user/s and an LLM and/or VLM
- a new WIKI for the tutorials
- super resolution model & denoising model support to (remove) noise artifacts or adversarial attacks from images the user may want to train with; i.e. mitigating some of the effects of Nightshade & Glaze
---------
- new tag, image, & caption statistics and visualization tools for the user to use from a data scientist perspective on how to best choose their data and augment their data
- a custom trained vision classifier on images with [nightshade, glaze, both, nothing]. To allow the user to know which data has been poisoned by artists etc. and if they need to be de-noised / upscaled on to mitigate the effects to some extent
---------
- tagging/captioning model/s will be downloaded on the fly if the option on the dropdown menu is selected, instead of having to go to the download tab to grab it beforehand.