This is a collection of custom styles I've created with carefully hand-curated datasets. Training is done locally on a 3080ti 12gb.
Unless otherwise noted, datasets used are a combination of my own illustrations and generations, artworks I've commissioned from various artists, as well as downloaded artworks made by various artists that I've heavily edited and postprocessed to fill out missing concepts or compositions where necessary. In the case of downloaded artworks used, I will try to include all proper credit.
If you are an artist and I've cited your name and you do not want your work used for training LoRAs/finetunes, please reach out to me and I will immediately remove the model from the site, remove your work from all training datasets I have, as well as add you to a blacklist to avoid using your work in the future.
These models are not intended for personal financial gain. Licensing is fairly unrestricted, although I ask that none of these models be sold without appreciable modification.
Additionally, these models are not trained for NSFW although they'll likely do fine for that given they are Pony-based. While I don't generally mind NSFW, I will remove images posted to my galleries if I deem them overly distasteful or offensive.
I will never release models in Early Access. Buzz donations are appreciated but unnecessary and will generally just be paid forward by me to other creators.
I use ComfyUI, but I've made attempts to include proper A1111 formatted metadata in my example images. All images will have Comfy workflows embedded, however I use a lot of custom nodes I've developed for myself so they may not always be useable for you; they should always have generation data viewable in a pythongosssss Show Text node somewhere though. For XY plot images, I will try to use only publicly available nodes from popular node creators (aside from my own SaveImage nodes for metadata purposes).
Please post your creations here, I'd love to see what you make!
Tools:
Training: Kohya-ss (https://github.com/kohya-ss/sd-scripts) with bmaltais' GUI (https://github.com/bmaltais/kohya_ss) / OneTrainer (https://github.com/Nerogar/OneTrainer)
Generation: ComfyUI (https://www.comfy.org) (https://github.com/comfyanonymous/ComfyUI)
Dataset organization and manual tagging: Nenotriple's img-txt_viewer (https://github.com/Nenotriple/img-txt_viewer)
Automatic tagging: SmilingWolf's WD tagging models (https://huggingface.co/SmilingWolf)
Image editing: Krita (https://github.com/KDE/krita) / Aseprite (https://github.com/aseprite/aseprite)
Base Checkpoints: AstraliteHeart's Pony Diffusion V6 (https://civarchive.com/models/257749?modelVersionId=290640) / Autismix_anon's AutismMix (https://civarchive.com/models/288584?modelVersionId=324619)
Additional Generation Checkpoints: ClamJam's ClamPDXL v2 (https://civarchive.com/models/361877?modelVersionId=521458)
Description
Third-party artists' works included in training dataset:
Totonero https://x.com/totoneroNNN
Polilla https://x.com/planetapolilla
Massakasama-san https://x.com/ma_sakasama
Flat(ish) colors and angular features. First step toward a personalized style for fantasy/science fiction settings with an emphasis on preserving base model composition. Limited by a small dataset and pure LoRA training, I'll switch to finetuned model LoRA extraction when I'm happy with where the dataset is at.
Training model is PonyV6, but intended for use with ClamPDXL_v2 as a base. Will generally work fine with other drawn/semi-real models, but I haven't tested at all with realistic models.
No trigger words. Training seems to have latched onto one specific image in the dataset, so "elf" and "pointy ears" tags will generally produce a specific style of ear.