On author's request, the model will be deleted.<hr />25/10/21 V2Slightly better prompt following, contrast and artist's style. What I did? Removed non 1024x1024 images. Turns out it is extremely critical for pixel art to maintain the training size. Which was obvious but normal images do not have such problems, mostly. Also some better prompts for dataset images since my "skill" improved since then and tags had some issues.<hr />25/10/06 V1It was risky to train model on 20 images only, but result seems fine. The model is Pencil Illustrous, 20ep x 16 repeats. Most of the images were 1024x1024, but some are 768x768 or portrait/landscape resolutions with x8 pixel scale. Don't really know if the difference in size impacts more than pixel scale for pixel LoRAs explicitly.

Files with SHA256: c2e3f3def06beb79ba112726267f772443c2473cd3828c275d5029a5b2f2a87d