So, this is completely stupid LoRA born from an experiment. I still like it, i'll experiment with it, so... i release it :D

The initial idea was:

• Using a Pony LoRA with Illustrious don't always yield good result due to the excessive anime style baked in.

• What if i could give just a bit of Pony UNET to an Illustrious checkpoint on the fly to help?

• What does "Pony minus Illustrious" equal to?

So, i started Supermerger and tried to make a LoRA by doing "Pony minus Illustrious"... and got a splendid Python crash due to the difference in CLIP. Not an issue, that's usual AND i wanted to get only the UNET diff, so, let's kick our good old Python up and do:

>>> from safetensors.torch import load_file, save_file
>>> illus = load_file("models/Stable-diffusion/base/illus.safetensors")
>>> pony = load_file("models/Stable-diffusion/base/pony.safetensors")
>>> len(illus)
2514
>>> len(pony)
2515
>>> new = {}
>>> for k in pony:
...   if "conditioner" in k:
...     new[k] = pony[k]
...   else:
...     new[k] = illus[k]
...
>>> save_file(new, "models/Stable-diffusion/base/illusTemp.safetensors")

For the more curious, the missing key in Illustrious is:

conditioner.embedders.0.transformer.text_model.embeddings.position_ids

This way, i got my 128 dimensions LoRA at full strength of the difference. I resized it to 32 dimensions and applied a 0.2 multiplier to "limit" the effect.

And here is the result in the gallery :D

PS: the cover image was generated using Regional Prompter + ControlNet to get the LoRA effect on only one side (approximately), with the help of the Split Screen LoRA.

NB: At higher strength, it start doing REALLY strange things, could be fun for experiments too :D