Trigger words:
Make the image pixel art styleDescription
First version trained on 20 images.
Could be better but I was happy enough with this result to publish a Version 1.
Works better on some images more than others.
FAQ
Comments (7)
how train a lora?any guide
ostris ai-toolkit
For a style lora, you need an image pair dataset, there are two ways to create them. One is to take real images and convert them to the style of your choice with loras, controlnets, etc. That way your output image will be synthetic. Difficult to do but this is what I did in this instance.
The other way is to find images that are already in your style, convert them to real photography with Kontext, and train on the reversed pair, so your inputs are synthetic. This is probably better but I've yet to try it.
Training on 1024x1024 takes at least 30GB VRAM and 20s/it speed so be prepared to be patient.
@n1eze Awesome lora ! could you be more specific about the pairs dataset ? do you name and caption them in a certain way ? any other tip ?
@Ggrue Here's a video tutorial by the ai-toolkit's author:
https://www.youtube.com/watch?v=WSWubJ4eFqI
Can you clone the effects from the Face app, please?
@nymical Thanks
@Ggrue you don't caption them in this case, each caption was just "Make this image pixel art style"
The pairs are just what you'd expect, one image in photorealistic style, and the other image in pixel art style - same subject composition etc. This is why it's difficult making such a dataset, they don't exist in the wild so you'll have to go synthetic.