If you liked the model, <a target="_blank" rel="ugc" href="https://x.com/sakuemonq">support the artist!</a> <hr />26/01/04 V3Some hyperparameters are changed, trigger word added. This looks like the most creative and stable version but I still won't use this in the merged dataset model.<hr />25/12/22 V2I have deleted some questionable quality images(artstyle- and details-wise) and removed the tags like solo, full body, character name. Despite it is logical to apply them, this is the style model in the first place. In my opinion the quality of generations is improved in result.<hr />25/12/12 V129 images of different characters on base IluXL. I also do not admire results with this one, but some generations with the previous model showed that on different config many models can be good.I also have a suspicion that placing characters not in the center helps model a lot with quality. In other cases there's a lot of noise in the center.I like how this model makes creative details like bizzare hair, magic weapons and so on. But probably it won't get in the big dataset I'm preparing for a while.The model best works with checkpoints I recommended in resources from the ones I have locally tested.

Files with SHA256: e8bca3e65edf64673923a33a1f97101d3782bda20b3d7856e4a9f00809be0a27