Progress for next model
I Have gotten the 2026 data, and images. The meta I am using is a little dated but will by syncing the old keywords to the new keywords.
I will be breaking the tags up in sections now like general: meta:
I am about 20% done doing my ai reviewed & human verified of the general tags.
I have generated about 12,000 images to help de-noise high issue tags. These images will all need manual review and human tagging to high accuracy as they will be used as a final fine tune pass to help shift out noise.
Current corrections is 500K. Mostly additional tags as removes from the first one for the most part clean most low hanging fruit out.
Overall cleaned data sits at approx 1.8M tags and rising.
Help is always appreciated.
My next model I will also be reducing the parameter count and increasing the dataset size.
Likely like this one, even with the reduced parameters I will hit a wall before training is complete.
An update for what is coming
Currently at 2 million corrections and climbing. These are AI assisted and human reviewed. 1.4M released there are some errors.
I will be updating the base dataset to 2026 for my next training than apply my corrections.
A golden dataset has been made for the following and climbing
ascot · bow · bowtie · bra · hair_bow · hair_ribbon · hairband · micro_shorts · neckerchief · necktie · panties · ribbon · scarf · short_shorts
hair_color (13): aqua_hair, black_hair, blonde_hair, blue_hair, brown_hair, green_hair, grey_hair, light_brown_hair, orange_hair, pink_hair, purple_hair, red_hair, white_hair
hair_length (7): very_short_hair, short_hair, medium_hair, long_hair, very_long_hair, absurdly_long_hair, short_hair_with_long_locks
hairstyle (9): braided_ponytail, high_ponytail, low_ponytail, short_ponytail, side_ponytail, wide_ponytail, sidelocks, long_sidelocks, very_long_sidelocks
multicolor_hair (6): colored_inner_hair, colored_tips, gradient_hair, multicolored_hair, streaked_hair, two-tone_hair
eye_color (12): aqua_eyes, black_eyes, blue_eyes, brown_eyes, green_eyes, grey_eyes, orange_eyes, pink_eyes, purple_eyes, red_eyes, white_eyes, yellow_eyes
multicolor_eye (3): gradient_eyes, heterochromia, multicolored_eyes
breast (6): flat_chest, small_breasts, medium_breasts, large_breasts, huge_breasts, gigantic_breasts
skin_color (2): blue_skin, purple_skin
penis (1): horse_penis
Golden data will be use to fine tune a model. This model will only be used to provide human assistance. So far about 200$ has been spent on the golden data. This involves sending guard rails and wiki definitions to Gemini. The fine tuned model will not be used as a full release model. Just aid with removing noise from datasets to further improve the vit. Corrections are also being supplied to people who train stable diffusion models
Penis sizes per danbooru's definition would be appreciated. Recommendations on what to further improve on is also recommend. IE someone was saying diapers often get confused with panties and shorts in diffusion models. So to improve this, I generated panties, shorts and diapers to use for the fine tune.
the next model will be trained on 6.2M images compared to the 5.2M from the first.
Fast update,
Currently general tags have gotten to 47% reviewed.
Golden data has expanded to 55K images. likely to be adjusted more.