What is this?
Trained on cute drawings of furries + animals + edited/upscaled anime frames
Even though this was trained on the same sort of stuff as the Illustrious version of uwustyle, the style is going to be a little bit different and more amateur-ish (this is what I intended with the original IL version in the first place but I could never get it to train quite how I wanted)
How to reproduce my results:
Prompt with primarily E621 tags and natural language wherever you want to be more specific.
Every sample image is first generated at 512 px resolutions, then upscaled to 1024 px resolutions and img2img at 0.5-0.7 denoise to improve the image, then upscaled again to 8.28 MP using lanczos and img2img with uwumerge checkpoint at 0.04 denoise to clean up upscaling artifacts/noise, then finally downscale to 2.07 MP (half size) using area to smooth the image out :)
Description
Combined the training datasets of v1 and v2 and then reduced down to the 86 best looking images, as I am aiming for a look somewhere between the two. I also added Illustrious-style aesthetic quality tags to approximate half of the training data (the better half).
Changed training parameters:
Batch size: 1 -> 2
UNet LR: 0.25 -> 0.5 (prodigy)
Dim/Alpha: 48/48 -> 64/64
Steps: 3600 -> 8600