All images showcased where generated using Pixai’s in-house model Haruka.
This was trained on a dataset of 19 specifically selected images using Pixai’s Lora trainer. The images selected all feature the slightly looser shading style that mochirong sometimes uses. The goal was to more accurately recreate mochirong’s style as I was unsatisfied with the results from the Illustrious dataset.
This was originally meant to be just a test for the dataset’s potential however It turned out way better than I was expecting. So considering I have no idea how the Pixai Lora trainer works I will most likely look to refine it in the future on a proper training service.
Please feel free to share any images you create as I would love to see what this thing is capable of.
Description
FAQ
Comments (4)
was it trained on pixai?
Yeh. Sometimes when I want to test a dataset without committing to spending buzz I’ll use the Pixai lora trainer as you get a couple free training runs. Unfortunately it does mean that I have no idea what the parameters are or even what base model Haruka is meant to be and sometimes it hardly captures the style/concept but it’s ‘free’ so whatever. I think It turned out okay on this occasion though
@Bert5535 how many images you used? because you managed to get a lot of epochs/steps, usually whem i use 100 images i get only arround 18 epochs and very low steps.
@MeltyGarnet I only used 19 specifically selected images. I don’t know how it managed 94 epochs considering it took like 12 minutes to train. Then again I have absolutely no clue how Pixai’s trainer even works in the first place so who knows.