~ Please upload your creations and give it a Thumbs Up if you liked it. It means a lot! ~
Please donate BUZZ if you want to see it on the generator.
A NSFW SDXL-Illustrious model that produces photoreal images with much greater flexibility and scene composition prompt adherence than any other photo/real models I tested so far, including realistic pony models.
ADetailer to fix faces is needed at times (not always), especially for people in the distance
Hera
Based on my finetune of Illustrious v1 and a number of realistic loras. It delivers almost photorealistic images and retains booru tags prompt understanding of Illustrious.
Use Eauler A for more creative output that sticks to the prompt, DPM++ SDE for more realistic look.
Photoreal - V0.2
Long time in the making. Hybrid of iLove Aphrodite and Big Love with sprinkle of custom Loras. CLIP 100% from Aphrodite.
Weaker character knowledge than Artemis & Aphrodite, however, look and feel of real photography. Good variety of faces.
Artemis
More realistic than Aphrodite with almost the same level of prompt adherence. Better faces, better lighting.
Aphrodite
Less real than the V0.1, however, it follows prompt like a book. Sometimes quality tags (masterpiece, best quality) help to improve prompt adherence, sometimes not.
All provided images are with the same seed (1), no cherry picking.
Use DPM++ SDE with Karras for best image quality & prompt adherence.
I haven't observed any glitches in my tests yet.
Use Adetailer for faces in distance.
Photoreal - V0.1
It is by no means perfect. Sometimes it produces glitchy results and artifacts. It requires ADetailer to fix faces at times (not always), especially for people in the distance.
However, when it hits, it hits perfectly. And it is very naughty. :D
This model is a result of nearly 3 months of testing various Dare & Ties merging technics, where more than 1TB of test data has been generated so far.
I wanted to create a model that understands booru tags, knows (some) concepts provided by amazing illustrious XL and produces photoreal images. In other words, I wanted a NSFW SDXL model that would provide greater adherence to complex prompts from scene composition perspective.
The complex multistep (~5 steps) MBW Dare & Ties merge included the following models:
Illustrious-XL - Base model for composition
Big Love XL - Base model for photo-realism
Nova Anime XL Illustrious - Reinforcing composition
Enormous respect to the models creators.
Parameters
For realism: DPM++ (2M, SDE, etc.), Karras, Exponential
For a smoother images: Euler A,
CFG: 4 - 7
Prompts
Keep it simple if possible. Excessive quality tags ruin composition adherence. Only put necessary tags in the Negative prompt.
Description
My first ever model published!
Parameters
For realism: DPM++ (2M, SDE, etc.), Karras, Exponential
For a smoother images: Euler A,
CFG: 4 - 7
Prompts
Keep it simple if possible. Excessive quality tags ruin composition adherence. Only put necessary tags in the Negative prompt.
FAQ
Comments (26)
By the way, if anyone has got a good resource with detailed explanation of how to tag images (uncensored) and how to train illustrious model, I would be VERY grateful!
This model is incredible! It's so good with realism and detail and composition, just wow! A bit of constructive criticism; I will say though one thing I did notice is that with cunnilingus it often times still tried to make it fellatio, like just having the tongue be like a long stick often, lol. Thanks a lot for this, so I want to help if I can. What I've used is https://github.com/kohya-ss/sd-scripts. It can extract loras from models, merge loras, resize them, tag images, and even training/finetuning a model. I highly recommend looking into it. Most of it already has the settings necessary for tagging, can tag one image or even a whole folder of them. Me personally I use; WD14 captioning with SmilingWolf/wd-v1-4-swinv2-tagger-v2 as the tagger. I haven't used realistic, mainly anime but the tagger should still work fine. I set batch size to 1. Ah, though as for training/finetuning full models I have not delved into that as my gpu doesn't have enough vram. When you do tag your dataset you want a program where you can then look over the images and edit the tags and/or fix or add any tags I recommend Boor dataset tag manager. https://github.com/starik222/BooruDatasetTagManager. Can easily tag and remove tags from a large amount of images at once or image by image as well. Really hope this is what you're looking for!
@Light7799 Thank you greatly for the feedback and the advice regarding training! Leave the model a positive review as well, please. If it's not a big trouble!
@wind11 No trouble at all, I had already done so in fact. Happy to help!
@wind11 Can I ask you about the "MBW Dare & Ties merge"?
@iamddtla Sure! I used Model Mixer extension for Automatic1111. There a few merging mechanisms available. Dare and Ties are among them. I have no idea about the underlining math, however, my tests showed that, generally, Ties merges retain more from the model A in terms of composition, and Dare merges retain more from model B in terms of Look & Feel (shapes, forms, skin, faces, etc.) if done with the same weight blocks.
@wind11 Is it this tool? https://github.com/wkpark/sd-webui-model-mixer
@iamddtla Yep. I used the hell out of it running with thousands of various MBW testing my prompts through XYZ plot. With different models and technics combinations
@wind11 Thanks for your answers!
@iamddtla All good, happy to help!
You could perhaps try replacing the base trained model with noob-ai as I think it is trained a little deeper and has a deeper understanding of the booru terms and roles.
I have never managed to get noob working at Automatic1111. I will have a look again, thank you!
@wind11 You should use the Epsilon-pred version, not the v-pred version.
@dawn66666666 Thanks, will check it out!
@dawn66666666 Is this the right version of Noob-ai? NoobAI-XL (NAI-XL) - Epsilon-pred 1.1-Version | Stable Diffusion XL Checkpoint | Civitai
@wind11 Yes, I believe that's the one. It's the same one I'm using anwyay
@Light7799 Perfect. I started preliminary tests. Will see how it goes with my merging method.
@wind11
I think 1.1 is very deep trained, I tried to train a realistic model with it, but the faces didn't seem to fit.
I don't know if it is a good model to merge.
But if you have problems, you can try an earlier version, like 0.5.
Of course, it would be even better if we could succeed directly on version 1.1.
@dawn66666666 Yeah, so far merges with BigLoveXL have not turned up very well - scene composition adherence is not as strong and the faces are distorted due to OUT00 blocks not sticking nicely. I will go through a few other real SDXLs and then go down the noob versions. Will see how it goes. Takes time as I go through thousands of iterations with various MBW to test scene composition prompts and faces to align.
There are already plenty of "real" illustrious merges with week composition prompt adherence that can't produce anything beyond solo posing.
@dawn66666666 I've done it :) The NAI based model is out:
NAILove - NoobAI Real SDXL Merge - v0.1 | Illustrious Checkpoint | Civitai
Interesting result!
Quite realistic and completely different style than Cyber Realistic_IL or Babes_Yogi_IL (these both look pretty similar) but I get way more deformed limbs or other problems with your model. Please continue your work, this looks really promising!
Thank you! The models you mentioned are trained models and have more control over the image. iLove is a merge with inherited capabilities and limitations of its parents. I am working to have a similar merge out of noobai and after this will work on finetuning the models.
Your checkpoint is very good, thanks.
Thank you!
Good news. I am closing in on creating a real NoobAI merge. I have a few successful merges already with various proxies stabilising composition prompt adherence. I am struggling to make a choice which one I want to publish :D
Yes its glitchy at times, but it's worth it!
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.