Generate sprites in the style of Final Fantasy Tactic Advance
Description
FAQ
Comments (6)
Seems to work fine for tag only characters on base Illustrious with square aspect ratio. Doesn't really work with character LoRAs or with checkpoints, though getting this to work at all is pretty impressive. Pics should appear within the hour.
Without knowing your training data/settings, the only advice I can give off-hand for a V2 is to 1: reduce the negative space by increasing the sprite's size while still meeting the 256x256 data minimum with next nearest in increments of power of two. The extensions to fix generated sprites all rely on downscaling anyways. 2: Use "pixel art" as a tag and use a trigger. Should make it a bit more compatible with others.
Edit: Don't know how applicable these are, but offering them in-case any are useful
1: I'm told pixel art wants minimal (no?) noise setting in training.
2: Marche's "ahoge" should be tagged as it comes up
3: If you haven't already, "moogle" and (especially) "Vierra" are recognized by Illustrious so you can use their sprites for more data with those tags. Same for "no humans" and monsters. using the back sprites "from behind, facing away" will double your possible data if not already used.
4: "Isometric" as a tag might help, I'm not sure though.
Thank you very much, this is still very experimental, my first attempt was so bad i'm very pleased with this one.
- I don't even remember how much noise i used.
- I made auto tagging, so there is a lot of improvement to be made with manual tagging.
- Dataset is 46 images, all front view with white background, i'm afraid of how adding variety and quantity will affect the result.
- I'm also very afraid of how adding the "isometric" tag will affect the result, isometric works badly with illustrious when used doing generation.
@davidconversano Don't be afraid of variety for styles (edit: when the distinction is clear in the tags). They want variety, and quantity alone can't make up for it.
You absolutely shouldn't be afraid to experiment and maybe end up with a junk LoRA. I'm only as good at this as I am because I was willing to try things and data types nobody else would touch (and documented what I did). (I will note however that styles are on the "deep end" of LoRAs and not what I'd recommend as a first project).
@NanashiAnon forgot to say that the original sprites are on a 64x64 image with a lot of blank pixels, and i then made the image 1024x1024 to train the lora, i don't know if it was the right thing to do
@NanashiAnon I started with a style lora because my biggest frustration with SDXL is the inability to create isometric sprites. But i'm sure i could find easier project to train with like clothes or character LORAs
@davidconversano Upscaling pixel art with next nearest is fine (edit: Though I recommend keeping to powers of two: 200%, 400%, 800%, 1600%, 3200%, 6400% size etc. since that's cleanest), especially when you want the results to be pixel art in scaling. It's other types of upscaling that end poorly.