Convnext Model:
https://pixeldrain.com/u/cnhn8ebK

Patched WD Tagger:
https://pixeldrain.com/u/2BNLUKGq

So deepdanbooru has become a pain to train and work with, so I decided to go have a chat with the person who did the tagger for Waifu Diffusion, and with their direct help now I can get us a model at that level. I have attached a current example comparison of current DDB vs 3 epochs on 1/4th of the images for Convnext. For reference, the source image was this: https://www.artstation.com/artwork/NGyeNg (sorry greg, I just happened to have you on my harddrive). Take note of the fact that while the tags are mostly the same, the confidence is a lot different and the tags missing from Convnext are actually more accurate, which is a really good sign.

First release of the model: https://pixeldrain.com/u/cnhn8ebK
Using it with WD14-tagger script, for now you need to replace it with my modified version (below this post). Overwrite yours/place it at "A1111-Folder\stable-diffusion-webui\extensions"
Model goes in "stable-diffusion-webui\models\TaggerOnnx"
At the end it should have this kind of structure: "stable-diffusion-webui\models\TaggerOnnx\Z3D-E621-Convnext\Z3D-E621-Convnext.onnx | tags-selected.csv"

Edit
Pub: 23 Feb 2023 16:32 UTC
Views: 538