This is my first attempt at creating a realistic model in Illustrious. After receiving a master class from QH96 on how it works and the differences with Pony, I decided to venture into Illustrious. I immediately began to notice the differences, mainly in understanding the prompt and many qualities and I fell in love immediately, I spent countless hours trying to perfect a model that would adapt to my tastes with what I had experienced in Pony and finally arrived at a result that I charm and I found it worth sharing, you will surely find some flaws since I declare myself not an expert but I hope you like it as much as I do, it is important to follow the parameters since it has many limitations, especially in terms of samplers, the parameters that I used are :
The VAE is not baked, use SDXL_vae
Upscaler: Im using Latent(antialiased) upscale 1.5 Steps 6 denois 0.35 Hires CFG 1.2, sometimes i use Remacri_original, less realistic but with cleaner images
Width and height: 896*1152
CFG Scale: 1.2
Steps: 8
The sampling method i use is Euler A-SGM Uniform
I don't like to put my models in early access because I like everyone to have access from the moment I publish it but any tips and Buzz are welcome.
Description
In this second version I tried to create slightly darker environments, improve the fantastic creation and give it a slight extra touch of realism, I hope you like it as much as version one, as I sometimes say, it is not necessarily better than the previous version, it is different and Everyone will have their favorite, for the moment I like it the most.
The VAE is not baked, use SDXL_vae
Upscaler: Im using Latent(antialiased) upscale 1.5 Steps 6 denois 0.35 Hires CFG 1.2, sometimes i use Remacri_original, less realistic but with cleaner images
Width and height: 896*1152
CFG Scale: 1.2
Steps: 8
The sampling method i use is Euler A-SGM Uniform
I like to use the following negative prompts although they are not strictly necessary:
glitched, distorted, blurry, low res, low quality, bad quality, worst quality, low detail, worst detail, cropped, censor, censored, text, signature, watermark, username, artist name, ugly, bad eyes, closed eyes, deformed, disfigured, bad anatomy, bad hands, monochrome, grayscale, sketch, oldest, simple background.
FAQ
Comments (15)
Thank you for your work first of all! 馃檹
The previews all seem to have some level of residual noise artifacts that commonly occur with acceleration LoRAs when dialed in too high.
Do you have a non-accelerated version you could share for comparison?
I think the noise you are talking about its because the sampler of the hi res, it's an aesthetic I like, maybe changing to a sampler like Remacri you get better results, I didn't develope the version without the acceleration but maybe later I'm gonna work on that, thanks for your feedback
Had a lot of trouble with these two. Couldn't follow prompts at low CFG and doesn't turn out well at high CFG. Maybe I'm doing something wrong.
this model easily produces fantastic details with minimal prompting
doesn't require any 'quality' keywords - it just produces realistic images
it's enough to prompt just a location to get a complete cohesive scene, for example 'mall, winter, outside' (yep, only three words) will result in a scene of bustling city life with people walking, mingling, wearing warm clothes and enjoying warm drinks
model is versatile - it can work with 8steps <3cfg for simpler prompts and regular 20+steps >5cfg for more complex prompts. resulting oversaturation is fixed with upscaling
realism came at the cost of fictional characters - they look like cosplay, also prompting several characters results in them having very similar faces and bodies
there's also prompt bleeding usual for XL models
some results produce too much cluttering objects - cafe doesn't need line of 7 teakettles :P
overall it's a solid realism checkpoint on it's own, but also good for upscaling from less realistic models and simply great for details - it certainly deserves more attention than it currently gets
It is really hard to use Sat Ill v2.
Guys keep CFG close to 1, use euler_a (or some other smooth sampler) with karras or sgm_uniform scheduler. Keep steps low near to 8.
PROs: if you hit the very narrow sweet spot it gives a clean, quite sharp, standard realistic picture with usual colors.
CONs: it lost the cinematic qualities of SatPony; background composition looks more messy; creativity is lower than SatPony have; quite hard to keep params working (very fragile version).
Can anyone tell me why half the illustrious models on this site are giving 404 errors? I just noticed this after getting an update notification for Thrillustrious v7, but it said the resource doesn't exist when I clicked on it.
Truthfully, I haven't even actually started using illustrious models yet; cuz I got a lotta other things I'm workin on. But, it's a bit unsettling. Is it to do with Civitai screwing over the creators on this site to make more money, and so they're all bailing? Cuz I heard something about changes to the buzz system being a kick in the teeth to creators.
Edit: Turns out this site is just bunk in general (in terms of functionality). Recently it's been hard to even get the thumbnails/results loaded properly on model searches.
I've found myself coming back to this model more times than I'd like to admit. It does faces well, and especially when mixed with a frontend, it refines extremely well. Also fantastic for inpainting. Highly recommended.
Definitely I have to make another version, I love it to
Beautiful Model with excellent performance
I used to always bitch at people that whined about vaes not being baked in, until I mostly stopped using separate vaes cuz most people do bake them, and then tried to train on the pony sat model, and Onetrainer was being a pain, not letting me train with a vae loaded, so all around, I think it'd be better if the vae was baked.
Cuz otherwise, this is a great model, but having to toy with the vae and it making training loras for it more of a pain, does mean I use this model a lot less than I would otherwise.
It's like disc games for console, who wants to get up to change the disc when ya can just click on a different digital game and play right away.
It's easy to bake a Vae for yourself in automatic 1111 or forge, o forgot to ro put VAE in My models but in the future i'll try to always bake it
@Sateluco聽I use comfyui. I've gotten pretty set on using it. I initially started back around July-ish with easydiffusion, then went to A1111, then SDnext, then in September when I upgraded and got a PC, I decided to give comfyui a shot, and been using it ever since.
That said, what kinda hardware does it take to bake a vae? Cuz I ain't currently a millionaire, so 3k+ for a single video card is a bit outta my price range.
So, I've got a 4060ti with 16gb. I figure it was the best bang for the buck, since I got it on sale for around 700, and anything better woulda costed me a grand or more, and for no improvement to the VRAM.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.