This was trained on a new dataset tagged with Google's Gemini API:

"View the images and provide all the details about it, including information about the subjects (mention the names of each characters if there are recognizable characters or logos), style (such as anime, manga, realistic, 3D render, pixel art, photo, drawing, etc.), details about a person or character (clothing, hair color, eyes, etc.), character expression, scenery, camera position and character view (front view, profile, etc.), concept, the scenario depicted in the image, and colors of various elements (hair, eyes, clothing, car, house, etc.). Describe the image visually in extreme details. Name all characters in the image."

I think this is the most stable version of this LORA. It was trained on 1000 images because that is the limit of CivitAI, I have a dataset of 8000 images tagged with Gemini.

I'm sharing a mini script created with GPT-4 for usign the Gemini API for tagging images.

https://mega.nz/file/Jx8U2JJJ#OV35ygry0GiJL4R7cuWxeNcmTk2IiFyrqi4kl1p6USo

How to use:

In the same place that the bard_tagger.py create a folder and name it "Dataset".
Inside of "Dataset" create another folder with the name "NSFW". All your rejected images will be moved here.
Move the images you want to images to the tag to "Dataset"
Open bard_tagger.py and replace "YOU_API_KEY_HERE" with your API.

Gemini is really good, it recognizes images very well and also recognizes many characters by name, for example, it recognizes Kurumi from Date a Live and Shinobu Oshino from Monogatari.

Dall-e 3 Lora