SamDoesArt Style XL LoRa

TLDR : Gen in 896 / 1152, strength : 1, end prompt with "by samdoesart"

Samdoesart was one of the first style I used back in december 22 on SD1.5, I though the other XL model already uploaded didn't do exactly justice to the material.

This Lora was trained using upscaled then downscaled back original images (90% of the original content with only sketches and low quality/old images removed).

All images were tagged using taggui (starting with blip2 then manually fixing all tags) in natural language. Everything was tagged (clothes, hairstyle, colors, backgrounds) so even though the original material has a narrow spectrum of subjects you should be able to prompt multiple things (From my tests it can even retain celebrity face similarity without using another Lora as seen in examples).

All training was done using 896 1152 AR, you should use that in your prompt if you are looking for the females face style. For everything else all AR should be applying the drawing style as it should but if you want the usual face proportion use that AR.

You'll get better results if you add "by samdoesart" at the end of near the end of your prompt.

No need to use the refiner (all samples images were just txt2img at 896/1152 with no upscaling/refining/afterdetailer etc). Comfy workflow / prompt was included in each.

Recommended negatives if those appears too much in your images : cat, dog, bird, pigeon, choker, grey background (they were tagged but can still appears randomly on small/free prompts).

The comb likes / dotted patterns are normal, they are brush the artist use in his drawing process. They didn't appear on SD1.5 models because of the downscaling. If you don't like them just add 3-5 steps of the refiner and it should remove them.

There is no NSFW content in the training set but it can somewhat do it through prompting using Lora's or another base model that can do it. Its harder to prompt in this case though expect random clothes appearing quite often.

The Lora is quite big (128 rank), I could try to make a smaller version later on.

Training parameters were : 128/1, repeat 10, epoch 15, batches of 8, Unet 8e-5, TE 2e-5.