Magic - Photoreal
Introducing MPR, my first attempt at achieving photorealism with the Illustrious base model.
The above samples were all generated with "Euler a" at 30 steps with a CFG of 5 at a native resolution of 1080x1440. No post processing effects or Lora's were used for the samples as I wanted to demonstrate the bare bones of the model and what you can expect from just basic settings and simple prompts. Too many people misrepresent what their models can do by using Lora's, embeddings, hires fix and other post process effects with their sample images and I wanted to avoid any such misrepresentation. Works well with most sampling methods and schedulers but based on the testing I've done so far I would recommend using either Euler a, dpm sde, or dpm 2s a with a cfg between 3 - 7.
now available for on-site generation!
ENJOY!
Description
FAQ
Comments (21)
I just used a run-of-the-mill prompt that I had lying around. What can I say. It worked. Just another day in the office, I guess.
wow, first results look very promising :)
What scheduler should be used with Euler A?
I usually set it to automatic
I'm so happy, I didn't accidently skip this one. So far it is my fav real ILL model, keeping balance between quality and flexibility + realism. Standing alone this one is a hit! Using it to refine with realistic dmd2 models is da sh1t!
I'm glad you like it. Rather than using it as a refiner with dmd2 models you may want to also try using the dmd2 lora with this model. I find it works quite well even when generating at 1080x1440. I usually keep the Lora between 0.6-0.8 to maintain some of the flexibility at 10-12 steps with a cfg of 1.1
Titan_Uranus I agree, this model alone is a blast. Thanks for the config recommendation :-) Going to try it now. Well, usually i use the dmd2 lora strengh 0.8, but you brought up 1.1 cfg, I almost forgot, that the +0.1 cfg could enable the negative conditioner. Good idea
Finger lickin Good!!
Great model
Are you sure this isnt "trained" model? very different from others. if you have included your own training data you should set the model to trained. 11/10 Looking forward to updates!
Nah, it's not. Quick test shows that it heavily based on CyberIllustrious v5 (~80% total) merged with something else (~20% total)
coyotte Cyberillustrious actually wasn't used at all in this merge. used over a dozen models but cyber was not one of them
not trained. merged over a dozen realism models into magic_ILL_V2 at low weights
Titan_Uranus interesting that with that method you've got a model, that gave me almost CyberIllustrious gens. On the other hand, it's these dozen realism models you've mentioned that ARE based on CyberIllustrious, probably, so it's expected xD
coyotte I think a lot of models share common bases, especially with Illustrious realism models. I tried to prioritize using uniquely trained realism models and merged them into my anime model at low weights to try and avoid introducing too much bias and maintain some of the knowledge and versatility of the anime model but it's a difficult balance to achieve. I find using Lizardon's profile useful to find unique looking merges and try to avoid using models that produced generic/similar results but I guess many others use a similar approach.
Incredible model. Thank you!
Makes great faces and for the first time in a while I don't have to put sliders all the way up so that it can actually make ADULTS by default. SMH checkpoint makers seek GOD. Anyway, my one complaint (2) is that it only does tiny nipples and that it doesn't get more complex prompts. Outside at that it good.
What do they even want? This isn’t some NSFW checkpoint made to suit their fetishes — and yet they're dropping pointless feedback like it's helping.
312312312312 I appreciate all feedback regardless. plz try not to discourage people from providing constructive criticism as it helps me to improve on future iterations of my model.
Just prompt for large nipples, large areola. It works
Exceptional quality and range for a photoreal model. Instant favorite.
Details
Files
Available On (3 platforms)
Same model published on other platforms. May have additional downloads or version variants.