My attempt in creating an SDXL Checkpoint merge that require almost no loras to achieve any desired output.
The main focus is to achieve a model that can give excellent results at CFG 1.
If you like my work you can buy me a coffee. Thanks馃挅
Main Features:
Works with a wide variety of prompts.
Extremely good prompt adherence especially with text.
Even at CFG 1 the text adherence is good.
Recommended Generation Parameters:
Sampler: LCM
Scheduler: Karras/Exponential/beta57
Steps: 14 -18
CFG: 1-2 (Don't recommend going above 1.8)
Models used for the merge mentioned in suggested resources.
Hope you guys like the model. Eager for the feedback..
Description
Merged it with few more of my favourite models (mentioned in suggested resources). After a few merges i ended up with two models with varying styles; v2 and v2(alt).
FAQ
Comments (50)
Stunning work!
Version 2 is phenomenal.
I can confirm that the best results were without the use of lora's. Thank you for the clear and precise instructions in the description.
I've only tried V2, what is the difference between V2 and V2_alt?
Well i did a few merges and ended up with these two. I couldn't decide which one to upload so uploaded both. v2_alt is pretty much same as v2 but a bit different composition and style.
And btw you can use loras with this checkpoint to get good results, it's just most of the time you don't need that's all.
Love this checkpoint! Was wondering if you ever tried to train a character lora for this? Any tips?
No i haven't tried training a lora with this checkpoint yet so don't have any tips as of now. Feel free to experiment yourself tho..
i tried to use lora trained on lustify , and its works even better than on lustify
Lora trained on original SDXL works fine. (for me)
Simiar to Lustify but 50% faster with same quality! great work man!
As I mentioned on your reddit post.... this is a masterful checkpoint!!! Rarely do I get super excited anymore, your model is top notch and I been enjoying it for hours now! I usually stick to my own but this is a keeper and will definitely get continued mileage from me!
Thanks for the appreciation.
idk from sampler LCM or from Scheduler:Beta but it rise my GPU card heat above 80 until it restart that danger sign !
this dmd2 plug-in so ? it's 4~6 steps then ?
best realism i have seen on sdxl. damn sdxl really has a lot more to give us. thanks for making this. awesome merge.
After looking at sample images, this reminded me of Epicxl and lustify, great samples, gonna try 'em out :)
I LOVE this model but I can't, for the life of me, force it to create a scene featuring an uncut D. I'm assuming there were none in the training data even though it's way more common across the world than cut. Sadly all the loras I've found for this suck.
Edit: F censorship
Really good. Yeah, as they mentioned before, it sort of looks like a Epicrealism but nsfw. Very interesting project. Thanks.
Instant classic! Ran many of my favourite prompts and results were instantly exceptional with recommended settings. Nicely done.
Why it sometimes gives oversaturated colors even on CFG=1?
Which settings are optimal for for hi-res fix?
the same setting should work.
i've tried so many checkpoint models, and this is by far the best one for realism imo!
Definitely gets a lot from epiCRealism XL.
Including the struggle of getting covered breasts. They are almost always exposed.
Any idea what the difference between v2 and v2_alt is?
V2_alt has less artifacts when Hi-Res Fix
Did you weave the DMD2 lora into the checkpoint or something? since its so clean at 14-18 steps.
And if so,what program did you use? Iv been messing around with DMD2 models and the lora has been a pain in the butt lol.
Yeah i used the DMD2 lora with very low weight and i used Forge webui for the merge.
@Kranos94聽I should try out Forge. Im still on A1111 but it doesnt have a merge option. (it has supermerger plugin but it has issues to install)
Does Forge have it built in? (or Reforge)
@KaiserWilhelm聽It's builtin
I need help training LORAs with this checkpoint.
So, I trained a character on the BigLove checkpoint, and I plugged it into a prompt using pikon checkpoint and the image was beautiful.
So then I decided to train a new character with pikon thinking that it would, obviously, perform even better with a pikon checkpoint. But then after I trained said character and ran it with pikon, the image quality was way worse :( .
It gave off that kind of washed out fuzzy image you get when you don't plug in a vae.
Because of this, I suspect the reason why the pikon-trained lora is bad has something to do vae or the dmd2 lora. Do I have to somehow "bake" the dmd2 lora into my character when I train it?
Do you know what I did wrong? I used the exact same images and parameters as the one I did when I trained the lora with BigLove.
Thanks!
It's probably because you used the parameters recommended for biglove. Since pikon has dmd2 lora merged in, loras trained on this should follow the same params recommended above( cfg1 and lcm).
Well i haven't trained any lora on this yet so can't say for sure. Feel free to experiment further.
@Kranos94聽Hopefully someone else will create a character lora from this and I can learn from them. Still facing the same issues even when I copy the same parameters.
This is an awesome model! Would love to see v2_Alt in an unaccelerated form (without baked DMD2) as well. Between the two, v2_Alt gives me much better results.
Glad you liked my work.
Also thanks so much for the tip.
Happy to give some Buzz for such amazing work. Please do consider releasing a non-DMD version of these models, the only issue I am encountering as-is, is a bit of same-face syndrome, which having a distilled lora embedded surely doesn't help prevent.
This model is great, any chance we'll get an inpaint version?
anyone have good setting for hi res fix?
Drecksteil geht nicht!
the best !!!!!
I鈥檓 feeling triggered by that naming 馃槄
Oh i'm sorry if i offended you but i swear the name for this checkpoint came into my mind completely out of the blue.
And your Legendary epiCRealism checkpoint is one of the reasons for creating this merge.
Thanks for your amazing work..
Tbh this is one of the best SDXL model out there, incredible work
How to eliminate open breasts in generations? 99% of portrait images are with exposed tits.
As god intended
Excellent sdxl model, but I have problem that its always showing a tattoo when I don't want it to
Put 'No tattoo' in the negative prompt
Can we get a viable scheduler_config.json for use with the Diffusers format? Perhaps upload to Huggingface?
This is an exellant SDXL merge, one of the best for NSFW content. The main problem with it, is that when you ask for a 1boy, men, man... you may get a woman (shemale...). But in general, the realism of that model, maybe mixed with a 2nd pass with bigloveXL1, is just fantastic! Thanks for that.
can't believe this is sdxl. I hope the low cfg won't hinder prompt adherence.
Any chance of a new iteration of this model? I love this one