If you liked the model - please support the actual artist: https://x.com/Resin_pixel

25/09/29 V1

The model was trained on 32 preprocessed images with 16x pixel size (1024x1024). All the images were jpeg so I needed to clean-up them and leave 8-16 colors. Most of the train images are witch girls, so that may affect the results. I found Illustrous Pencil model better with LoRAs so I switched from WAI and Plant Milk(thanks to Skormino for head-start!) training.

Obviously, the model does not fit for fancy images. But it handles prompts absent in the dataset quite well.

I found sometimes that weights 0.5-1 are good. Sometimes lower weights are better pixel art, some times it's even just a neat flat color. If you don't care about investing buzz/resources experimenting - fell free to share!

I haven't tried to test the tags, but they may be useful: pixel art, limited palette, {color} background, {color} outline.

Now, the model weight is too big. That was my safe spot for buzz saving, on previous models experience. But the results were good from the first epochs, so I plan to make V2. But the catch is for V2 I want to augment dataset with new images, based of the generations, and it's quite tedious, so you could contribute too with your posts.