Illustrious
Ideal resolution is 864x1152
Trained at 1024 with buckets at 1152x1536
Weights best at .9 - 1.0
CFG best at 3.5 - 5
Add "white border" to prompt to be more authentic to the originals.
Pony
The Illustrious version clearly handles pixel art better but it was requested.
Ideal resolution is 864x1152
Trained at 1024 with buckets at 1152x1536
Weights best at .75 - .9
CFG best at 5
Add "white border" to prompt to be more authentic to the originals.
SD15
Ideal weights appear to be .75 - .85.
Trained at 576x768.
With LoRAs: the weight might require lower values.
Description
FAQ
Comments (7)
That's really impressive, especially with the limited data and "pixel art" absent from the prompts and training data. What kind of trickery did you employ to make this work so well?
The dataset is precise, otherwise... luck? This was trained on Civit - so the defaults on training Illustrious models + a flip augmentation.
@UncleJert Interesting. My one pixelart based style wasn't one I was planning to redo for Illustrious, but I'll give it a try once I get everything else converted.
@NanashiAnon I forgot to mention. I didn't resample the image data. I used 'nearest' on upscaling to preserve the original style.
Neat
The Pony version of this came out pretty good in my opinion, well done :)
Arguably doubles as a Makoto Wata style LoRa, (j/k, but only half)