These are my Mergestein realistic experimental models, a lot of merges merged in a merge to make a merge. Works for anime/toon/game/real life characters and can do furry too.
on-site generation with "private" NSFW support:
Uncanny v2 (shows as uncanny-)
If you want to support me, just upload an image on the model page, it doesn't matter if is good or not, i receive blue buzz for LoRa training.
The images i upload on the model are generated using Euler Beta and without any LoRA, CFG Rescale, Hiresfix, ADetailer or Upscale. So, the quality can be improved. Like all illustrious models, close distance is perfect, medium is ok and far is ugly.
All models already have a CFG boost merged in the model (it works well with low CFG, use less than 5 CFG), so is not recommended to use any CFG LoRa, you will only get burned images and some detail boost LoRas can affect too, if you still want to use them, use low weights (less than 0.3).
Yes, all models work with NSFW and Age stages works well, so please don't use it for bad things.
Samplers, Schedulers and CFG Recommendations:
The less CFG, more realistic. The more CFG less realistic.
Euler A Normal/simple/SGM Uniform 2.5 CFG (max 5 CFG) if you don't like skin texture (Skin pores, wrinkles, etc)Euler Beta 2.5 CFG (max 4.5 CFG) for better 3d/realism, close to DPM but with much less "noise".DPM++ 2/3/4M SDE Karras 2 CFG (Max 4 CFG, more the CFG = Brighter/burned). More realistic but too noisy and adds too many details.For Euler/Euler A start with 16 steps max 30 steps.For DPM start with 20 steps max 35 steps.Resolution:
Works with all SDXL resolutions, i personally use 1024/1152x1536 and sometimes 1280x1536 (with this you can get body extension).Recommended general base prompts:
Positive:
best quality, masterpiece, excellent, high Resolution, aesthetic, photorealistic, photo backgroundNegative:
worst quality, bad quality, low quality, lowres, low resolution, anatomical nonsense, artistic error, bad anatomy, bad feet, bad hands, extra digits, conjoinedPrompts to get more realism and quality:
(it can reduce character accuracy, recommend using with LoRas or more weight for anime/toon/game characters)
Positive:
best quality, masterpiece, excellent, high Resolution, aesthetic, realistic, photorealistic, photo (medium), photo backgroundNegative:
(Remove "(toon (style), 3d)" for better character prompt follow and creativity)
worst quality, bad quality, low quality, lowres, low resolution, anatomical nonsense, artistic error, bad anatomy, bad feet, bad hands, extra digits, conjoined, (toon (style), 3D), traditional media,Credits to respective creators, i just merged things together (sorry, forgot the realism models and LoRas used on my personal realistic merge, it's old), for my Wai x Janku + NovaMoon-PM ILL XL all credits are available on the model page same as my Uncanny Animu Plus.
Added to Uncanny Realism dark by ntc
Uncanny v2 added Ahegao - Pleasureable Expression by 6__6, as a fix for eyes and mouth (and ahegao looks better too).
Uncanny-A Plus added to the merge PornMaster-Pro noob-v5, CyberRealistic XL - v8.0 a, cfg scale BOOST and NAI vpredfix (0.1 weight strength each one) and uses my Mergestein Animu Plus as base instead of vZero (and still has the aheago LoRa and other NSFW ones).
Description
Same name as the model due to being more realistic, more natural skin tones, darker lighting without affecting bright zones (i see that people like dark models) and more skin texture. Fingers sometimes still have problems.
I recommend adding the age stage of the character for better consistency (sometimes it can make them much older/younger than they are).
If you don't use any type of "Unrealistic character" it can do realism pretty good, but not at the level of my Realistic ILL model.
FAQ
Comments (18)
With the release of the NAI realism model, which one do you think is the best at realism between your ILL-based and the NAI-based? Thank you
For pure realism the ILL version is better, i like it more for that due to having a more natural lighting. The new version i uploaded is better for anime/toon/game characters and makes them more realistic than the uncanny version, but still not at the level of the ILL version.
@Patuwa Thank you, I'll give the ILL version a try then!
@Patuwa Oh. My. Gosh. I tried the ILL version, it's an amazing merge, hats down!
@Patuwa does the noob version work well with illustrious loras? Ive never used Noob
@Lexcorp_AI Pony / IL / NAI / WAI are all SDXL-based, so it's not optimal (as it's not trained on the ressource you're using) but it works.
@TheGlowingGuardian what is noob based on?
@Lexcorp_AI "NAI" is NoobAI, so SDXL as well
@Lexcorp_AI Yes, after all NoobAi is a fine-tune model of illustrious made specifically for anime, uses booru/pixiv database for better anime composition and characters without the use of LoRas. It works with Illustrious and some SDXL LoRas but due to how NoobAi was trained, some LoRas can produce bad results, so you need to test by yourself if the LoRa works correctly.
@Patuwa thank you!
Great realistic Illustrious model, well done!
Even though it's a bit overdramatic with the lighting, it's actually a very flexible and cool model!
I just merged it with a little bit of Place2Play to soften lines and lights, but otherwise, surprisingly good adherence and results. Will keep an eye on the evolution.
I'm already testing a Realistic and Uncanny model with a more natural and soft lighting, but most people like dark/dramatic lighting, that's why the lighting of the first models are like that. It will take some time to release the models cuz testing takes too much time on my low-end laptop.
@Patuwa loving R-v2. There's some great balance between realism and cuteness that makes it real but stylized enough so it doesn't hit the uncanny valley too hard.
[Realistic ILL version]: I like the level of details we can gen with the model, but it feels like it has a really really strong "sameface" effect. Any idea how to reduce this? Thank you
Yup, the "pretty generic face" is a problem, the only way to reduce it a little is by adding an ethnicity and age stage to positive prompt, using a LoRa of a character/person or try using this LoRa (use negative weight like -0.4-0.8): No More Sameface - Face Randomizer - SDXL v3 | Stable Diffusion XL LoRA | Civitai
I will try to merge with a model with more diversity on faces on the future to try to fix this problem.
@Patuwa I tried ethnicity changes, the sameface LoRA at -0.7 and even character LoRAs won't work. Even cartoon characters get that 'generic' face instead, with their haircut or clothes. The model is still impressive, really. I'll have fun in the meantime, and can't wait for future updates then 👍
This is a very fun checkpoint that I keep coming back to.
Details
Files
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.