TRAINING V3... =)

Yaoi Diffusion V2

768 resolution model finetuned on yaoi, bara, furry, s...., s.... c.., fine arts and reallife males, in short a general homoerotic model.
I've seen people finetuning LORAs of artists, so may as well upload a model that has lots of them, the list of tags can recognize is here for version 2

https://gist.github.com/iszotic/d7aad5c056601aaf7cdcc2e0fedcb2a4

HOW TO PROMPT:

[short description] by [artist], [e621/gelbooru tags separated by comma and space]

negative tags: _old, simple background
the old tag sometimes makes the style better, sometimes it doesn't, I only used bad-hands-5 for the examples but other negative embeddings work fine too. scrawny works better than emaciated for skinny characters, civitai censors that word LOL

ex:
An anthro furry dragon male laying on bed by artist_tag, 1boy, male focus, solo, pectorals, penis, realistic

Also check out the demo images with the prompts used, the high resolution was achieved using controlnet tile resampler with 2x resolution.

_other, is another style, normally is the cellshading version or a simplified style.

Use the booru convention, like use 2boys instead of male/male

MIXING STYLES

there are 2 ways to mix styles:

1) making the diffusion process exchange the artist tag in each step (auto1111)

by [artist1|artist2|artist3|artist4]

2) or using the tags all at the same time:

by artist 4 by artist 3 by artist 2 by artist 1

in 1) the first artist takes the lead, the features will resemble more of this artist, although the mixture is more noticeable. Works with Euler a, Euler and DDIM samplers the downside is the quality is not good.
in 2) the last artist takes the lead , the features will resemble more of this artist, the mixture sometimes is not effective. Works with any sampler, the quality is better.

COMERCIAL USE: it's ok as long as the resulting style doesn't resemble any specific style.

Features:

• over 900 artists tags of homoerotic artists (including myself, lol), no tags where used for pure 3D artists.

• NSFW and SFW

• some artists support more than one style

Training details:

• Trained from SD1.5 vanilla

• Dataset of 220K, epoch size of 95K, rated dataset of 3K, dropout probability of rated dataset 0.5, dropout probability of not rated dataset from 0.0 to 0.15, depending of aesthetic values from https://github.com/LAION-AI/aesthetic-predictor and https://huggingface.co/cafeai/cafe_aesthetic

• Images were sourced from booru sites, and tags were sorted using deepdanbooru, the e621 model comes from zach and the wd14 swing model, if images were not from booru sites the tags were predicted, the order of the tags were randomized 2.5% of the times. Also used blip2

• Everydreamertrainer2, gradient checkpointing disabled, and gradient accumulations.

• Training schedule: Epoch 0-17: Lr: 6e-6, B:4x64, Epoch 18: Lr: 3e-6 B:4x32, Epoch 19: Lr:1.5e-6 B:4x16, Epoch 20: Lr:7.5e-7 B:4x8, and 3 epochs of rated images. B = batch size

• Only a maximum of 100 images per artist were used in each epoch, if an artist had 500 only a different set of 100 was used.

• zero frequency noise ratio = 0.01