Older Description for v2 and v1 -
This checkpoint is made to retain most of the knowledge of base Illustrious/Noob while making everything as realistic as possible.
--Prompts have been attached to all of my gens--
I had one vision when I started working with Illustrious - creating a realistic merge that did not ruin the base model's extensive danbooru knowledge.
Prompts for the showcase images have been taken from all the best Illustrious model showcases and galleries. All showcase images for this model are one shots with random seeds and one hires fix. No Adetailers or inpainting has been used for those. In most cases, v2 feels way more real as the faces are much less anime-ish now. If you want the skin to be like v1, just use the suggested LoRA at 0.4 strength.
V1 has some problems with residual noise and general clarity. V2 should have more depth to its "realism", and should offer better compositions and quality in general as compared to the first version. Do try it out. (The first version can still be quite unique for "grainy" gens).
USAGE GUIDE:
Now, here's some tips for squeezing the most amount of realism from this checkpoint (these are just my observations, feel free to experiment):
Make sure that you use lower values for your CFG (3-5). CFG 3 and karras works incredibly well for realism.
Use DPM++ 2M with the "Beta" or "Karras" scheduler, and do a highres pass at 1.5x at 0.33 denoise and half the original steps. (Please experiment with different sampler and scheduler combinations if you can).
Remove all instances of artist names and other "painted" styles from your prompts (unless they tend towards realism).
For the positive prompts, include this in the first part (Kinda IMPORTANT):
Optional - "cinematic lighting"
real, 4K, RAW, photoreal, DSLR photo, real skin,For the negative prompts, maybe use this?
cropped, lowres, worst quality, sketch, unfinished, greyscale, faded, low saturation, jpeg artifacts, watermark, text, anime, CGI, 2D
Online run on Tensor -
(Generating onsite on Tensor gives the gens a grainy look (v1 only), making them even more real. Absolutely no idea why this happens.)
This model responds to prompts really well, and in a similar fashion to base Illustrious/NoobAI models. This also implies that you can use artist styles in your prompts for uncertain ways in which the realistic part gets affected.
v1 Related:
Now, some of you might not call this model very "realistic", but it certainly does have that 3D style that I really love. The end goal is always to achieve more realism, while making sure that the model's base knowledge is not disturbed.
There are other Illustrious based models that can provide you with more actual "realism" (debatable), but all of them have to sacrifice some of the model's incredible compositional and conceptual abilities. Some models lose character knowledge, some lose the character clothing knowledge, the dynamic compositions, the multi character abilities and much much more in search of truly realistic outputs.
And that is exactly what I hope to avoid with my merge, keeping almost everything in the model, while bringing the knowledge of the 3D world into it, and slowly increasing the realism.
If you want to support the creation of more models like this, or you just like what you see (I could REALLY use it): ko-fi (｡♥‿♥｡).
Also, no plan on charging early access for this one, so buzz donations are appreciated a lot too.
Models Merged (Credits and respects to the original authors of these models for allowing sharing merges) -
ReallyIllustrious (base) - TrueToLife_Fauxto
Alchemist Mix (Creative Realism) - Urameshi
coco-Illustrious-NoobAI-XL-Style - COCO_OIOI01
Illustrious_Sky-(rim) 3D - stableskynet171
PornMaster-Pro 色情大师- Illustrious - iamddtla
CyberIllustrious | CyberRealistic - Cyberdelia
TAME Pony: The Authenticity MachinE (v2) - zyxt99565
IllustriousSat-Real - Sateluco
Quintessence (Magnum Opus) - Realism | The Delicate Balance | PDXL - Me
ReallyBigLust - Illustrious Realistic - Hikarias
Description
New LoRAs merged along with a tiny bit of CyberIllustrious to fix some background issues. The model has also been normalized for more stability. Faces are now a bit closer to real life.
FAQ
Comments (11)
Very good, it does have very comprehensive knowledge, even some weird things, seeing the IL model perform like this makes people look forward to the future
Yea.. Cuz the future is full of 'weird things'.. I don't even mean that sarcastically, modern reality is a nightmare..
Generated very ugly faces with big nose.
Dislike.
Just add "ugly, huge nose" to your negatives and you'll get the faces that you desire to see. Try to think of the "ugly faces with big nose" as the default face for the model, which can be modified any way you like with simple prompts. This was essential in order to try and get rid of some "sameface" problems that most models suffer from.
@nottheninja Loras also work wonders with illustrious realism models, if you got the hardware and/or the patience to train some. I rarely use any models without Loras, well, for one, most models are so female-heavy that you can't even get consistent male outputs(without futas or engorged nipples) without at least one lora. But apart from that, loras are fun for mixing in non-existent characters. You can take cartoon/anime characters and make very real versions of them.
Oh, and img2img, which I also do often. A blend of loras and img2img can be great, cuz you can get it close to what you want, then drop that generation into an img2img workflow in comfy, and get it even closer to realism from there(exponential improvements :-D).
I don't think the face is ugly, but it's definitely the same face almost every time. It's at least a unique "same face", but it's always the same person.
@Jellai Prob cuz most images used in AI are of 'good looking' people (AKA: clones lacking in unique or asymmetrical features). So, what the model ends up putting out are typically generalizations of the 'good looking person'. hence it's unique, but repetitive. One thing I find more repetitive is hair.. Well, maybe not for women, but maybe. Cuz there's so many women in the training, but then, most women do have severe clone faces due to makeup, and most have long hair due to societal gender norms.
I get REALLY good results with DPM++ SDE - SGM Uniform - 30 steps - 3-5cfg. Using recommended positive prefix and negative prefix, but I added "ugly face, big nose" to the negative prefix.
excuse me ? what can Illustrious (or noobai) do when added to a realistic sdxl ? benefits ?
I mainly aim for the character knowledge and better prompt adherence of Illustrious, while trying to get a realistic look. SDXL is still better for just realistic portraits, but Illustrious can make better hands, poses, and generally has more variety.
Amazing
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.