Phoenix Pholly R

This model is the product of my attempts to graft some realism from a couple PDXL fine-tunes onto Illustrious. I was maybe 50% successful at that, if we're being generous, but it's still potentially interesting. It's definitely nowhere near as good as a proper fine-tune would be, but it's also something I could manage on my current GPU.

I've found it works best with "photorealistic" as the first tag, ~30 steps, and Euler a as the sampler. Too many steps make the image messy, other samplers are a bit hit or miss. UniPC seems to be more reliable than DPM++ 2M, though it's not available on CivitAI.

It seems to work fine with the normal SDXL VAE.

For anyone who's curious, the method I ended up using to merge Pony fine-tunes onto Illustrious was the "add difference" mode in A1111. I used Illustrious as model A, the chosen fine-tune as model B, and Pony v6 as model C. This worked so long as m was less than 0.4 (I settled on 0.3 for all four models I merged). I then did 50/50 weighted sum merges to collapse all four of the sub-models into one.

Directly merging Pony-based checkpoints onto Illustrious resulted in something that was less than the sum of its parts, in my experience.