My first attempt to create a photorealistic SDXL-Model.
I hope, you like it.
I merged it on base of the default SD-XL model with several different models. (Around 40 merges)
SD-XL VAE is embedded.
I strongly recommend ADetailer.
You should set "CFG Scale" to something around 4-5 to get the most realistic results.
v1.2 is an updated version with some added elements with a slight tendency to be a miss of the dark. 馃槉
v1.3 has new improved details, better nsfw results, I also refined the output for better photo realism, the missus is still here 馃槈
v1.4 has better results for NSFW images, I also refined some details, less missus but improvements in base model 馃槉
v1.5 is better fine tuned, improved details, refined output
v1.6 is an improved model with better details and better NSFW compatibility.
A music video, I made using this model and Udio and wav2.lip and KlingAi and much more is here:
Description
FAQ
Comments (23)
I'm blown away from version 1.3's results, absolutely amazing! - Can't wait to try the 1.4 version but have no clue how that one could even top the 1.3 version. - All of your hard work has paid off... Congratulations and many, many thanks for the models.
Cool that you like my model 馃槉
The most significant changes in v1.4 are better results in NSFW images.
If you are happy enough with v1.3, use it then 馃
It would be interesting which other models you did use for your model and honest to the creators to call there names of the model that you merged in馃槈馃憤
In fact, I merged the SDXL base model with trained Models, I made over the time, until I was happy with the result.
I mainly used the base images of my other SD15 photorealistic model and re-trained and merged them together. (After upscaling / regenerating)
@Devilmarkus聽nice I did the same with my SD 1.5 model data's nice work btw 馃槈馃憤
love it :)
deleted all my other models, only using this now...
One of the best models I've ever used.
It takes endless for me to create one pic, with all othter creators i can create a pix with 20 to 50 steps in seconds. What am i doing wrong?
Could you be more specific? I guess you used SD-15 models before, but this is a SDXL model.
I really dont know what the difference is. I mostly use epicphotogasm xplusplus (etc.) or cyberrealistic. What is the difference to SDXL?
@Crankerrr聽the difference is the base resolution, SD-15 models are using 512x512px and SDXL uses 1024x1024px as base.
For SDXL you need of course a graphics card with more GPU ram, and the rendering is slower than with a base SD-15 model.
@Devilmarkus聽Thank you! So if i reduce pic size by the half, it should be same fast as with SD15?
@Crankerrr聽if you reduce the pic size for a SDXL-model, you may get more worse results than with a SD-15 model, as it's base resolution is 1024x1024
@Devilmarkus聽Allright, its normal that a 25-50 steps genearation with a SD15 takes some seconds, and with SDXL it shows me1-2 minutes?
@Crankerrr聽depends on your graphics card, for SDXL you should have at least a RTX 4070 with 12gb GPU ram, better are 16gb GPU ram, my laptop for example has a RTX 3060 with 6gb only but with low gpu ram parameter, it renders SDXL, too, but within 4-5 minutes ;-) My PC is way faster
@Devilmarkus聽Yes, then i have to stay with SD-15 :) Thank you so much
@Crankerrr聽you can try my old photorealistic model for SD-15 here:
https://civitai.com/models/99548/devlishphotorealism
@Devilmarkus聽thanx bro, will try now
how do you even do this? i have a 4090 and can't even train SDXL
Did you try that with fp32 precision? Don't do so, use fp16...
One RTX is a bit weak but it should work?
so use Fp16 and the 4000 series don't use NVlink so idk how it would be possible to use more than 1
Well for me it works, veeeeeery slow, but it works :)
Btw: It also may help to train several LoRAs and merge then with kohya_ss to your main model...
So you have less GPU ram usage while training.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.