Question about training, especially about caption
Hello, this is a new organization who plan to train a new open-source model using large datasets and want to ask your experience about captioning.
- Did you use the original Danbooru tags or make some adjustments on the tags. If latter, how and aim what?
- Could you share some strategy about quality tags and functional tags?
- Are there any implementation method, such as tools for captioning, labelling quality tags?
- Did you change captions' order (manually or using tool)? If so, how and aim what?
I'd be appreciated if you could share more training experience with us. Thanks!
Hello, thank you for visiting us. We wish you success in training your model. Regarding the question you asked about captioning, it will be answered below. ^^
Yes, We utilize the original Danbooru tags.
Our strategy includes using a dataset-builder, which outlines our approach to curating from masterpiece to worst quality tags and and functional tags.
For captioning and labeling quality tags, we use the WD tagger as a tool for captioning and Aesthetic Scorer for labeling quality tags
Similar to the strategy for creating quality tags, to alter the tag sequence is also included in our dataset-builder where they are sorted starting from first_general_tag, tags_character, tags_copyright, to tags_artist, etc.