"SPIN-Diffusion generates images with superior visual appeal than previous fine-tuning methods"
It's sort of like DPO, since it has been trained on human-preferred aesthetic images. Or can be used like a detail enhancer.
I'm still experimenting with it what it can do, but have provided a few different sizes of LoRA to play with. You can try it with different weighting for variety, or just stick with 1.0.
FYI Clip weight LoRA parameter will have no effect, since this was trained on the U-Net model only
The original model can be found here: https://huggingface.co/UCLA-AGI/SPIN-Diffusion-iter3. I merged the U-Net weights with SD 1.5 base and extracted into some LoRAs of varying dimension. reprinted from: https://huggingface.co/otterpupp/spin-diffusion-v3-sd15/tree/main/loras
Description
dimension 128 lora
FAQ
Comments (3)
Not that I'd use this in the first place with all the restrictions on it, but the image doesn't really demonstrate clearly at all what, if anything, this is supposed to do. It's also huge, so it's an especially hard sell, particularly since you used ADetailer on the results and didn't just let them speak for themselves.
Or you could download this in good faith from the huggingface links, where its license is declared as Apache 2.0, and on-site generate your heart out.
I don't really understand what exactly these aesthetic slider loras do either, as I'd have to learn a lot more before I could comprehend a paper on quantifying aesthetics, but usually they up the saturation or reflections or make a character look more focused or interested, as with the puppy example, something that'd probably make your result more noticeable at least. Whether or not I like the effect on my stuff is a roll of the dice, but I'll be generating my share of thirsty ginger men with it myself to find out.
@hushicho I have updated to remove the accidental restrictions. Please it out and experiment?