This model represents the new Willy’s Realism line and is now considered the Primary branch, while the original Willy’s Noob Realism model is considered the Legacy branch.
!!! If you are having issues with this model, please scroll down to "Common Problems" before commenting. !!!
Version Information
Core v1
This model was rebuilt from the ground up to serve as a clean, consistent baseline for realistic image generation using a Noob-based foundation. While it uses many of the same modern resources and backbone models as previous versions, this version focuses on improving overall generation quality, and upscaling performance without the compounding issues caused by heavy merge inbreeding.
Prompting methods remain very similar to Willy’s Noob Realism, and users familiar with the Legacy branch models should feel right at home.
Because this is a fresh baseline, some of the more specific improvements or quirks from previous versions are not fully carried over yet. These will be reintroduced over time as the model continues to evolve.
Recommended Settings:
TL:DR - Prompt normally using the Noob quality tags. For realism add tags like "detailed, realistic, photorealistic, hyper realistic, sharp focus". My personal favorite generation settings are Euler Ancestral or DPM++ SDE with Beta but feel free to play around and find something you like. CFG range of 4-6 works best for most images.
Recommended prompting format:
<quality tags>, detailed, realistic, photorealistic, hyper realistic, sharp focus, <style tags>, <character tags>, <pose tags>, <background tags> Positive prompt example:
masterpiece, best quality, absurdres, newest, detailed, hyper realistic, (realistic:1.25), (photorealistic:1.25), (sharp focus:1.15), 1girl, anthro, wolf, grey fur, fully-clothed, standing, looking at viewer, posing, bedroom,Recommended negative prompt:
blurry background, low res, out of focus, (blurry:1.2), worst aesthetic, early, old, worst quality, anime, manga, cgi, illustration, digital artwork, 3d render, (painting, picture frame), watermark, text,If you are getting unwanted humans or human-like body parts add "human" to the negative.
These are my personal recommendations, but feel free to experiment and use what ever settings you feel works for you!
Common Problems:
My recommended settings are just my personal recommendations for use with ComfyUI and have not been properly tested with other front ends.
If you are seeing artifacts please make sure your generation settings are set correctly and that you are using the correct VAE. Reforge may use a bugged VAE by default.
This model uses V-Prediction and related settings may need to be set manually.
Some Loras trained on EPS models can cause issues with color, lighting, or other aspects of your image when used with this model. Not all Loras will have these problems.
Description
This is the first release version of Willy's Realism Core.
This new Willy's Realism variant was merged from the ground up with no inbreeding of older versions using a combination of old resources that were the backbone of the older versions of the model and newer resources while focusing on overall generation and upscale quality.
This model variation does not include all of the improvements from previous Willy's Realism models, but is the new baseline for the next generation of Willy's Realism models.
This is a v-prediction model.
FAQ
Comments (21)
Hi, do you have any plans to make a realistic furry / feral model based on Qwen or Z-image? Or something else in the 12-18 GB range...
For some reason, no one makes such models (except Ratatoskr, but I can't get decent results with them).
I made a Z-Image Lora that is here on Civit. It took 3 days to train and the dataset was partially flawed. It can get you some pretty decent results out of Z-Image but does have the 3d render sort of style that the Legacy Branch was basically known for.
With the right prompting though you can get some pretty decent results out of it. Just be warned that it doesn't play well with some other Z-Image Loras because it was trained before Base was out.
https://civitai.com/models/2280387/willys-z-image-anthro-lora
@willywizard It looks too 3D.
And I'd really like to see a larger model that can do a lot of things out of the box.
Maybe you know any such models?
@MrA1ex If you know how to prompt for it properly, I’ve seen some decent furry things coming out of Klein recently, I personally haven’t looked very into it though.
Lodestone’s Chroma model is pretty okay, it’s a furry fine tune of Flux 1 (Schnell I think? Although has unique architectural changes to it)
From what I understand though, what you are looking for doesn’t exactly exist right now, and I don’t think anyone is currently working on anything like that.
I personally would if I could, but I don’t have the hardware or the funds.
Chroma is super capable, having tested it recently. I wish more people would invest in it, it has great potential.
Either I'm doing something wrong, or the model doesn't work with img2img...
Its a V-Pred model. In my experience it works fine with img2img, but you are going to need to probably use a lower denoising strength and or lower the cfg depending on what problem you are having. Or maybe just use a different sampler/scheduler all together specifically for inpainting because some of them. And example is anything mixed with Karras denoises too much at the beginning because of how the math handles it and will almost always screw things up.
This is GREAT!
I don't think I can do justice in my praise, but I'm going to try.
I'm one of, apparently, a minority of people who create anthro pictures, who wants something that is strongly realistic, not faux-realistic. When I've spoken to others about what they use/like, the "realistic" suggestions I get look more like 80s anime characters. I've had people shit-talk my own work, because they don't like "too" realistic furry.
Because of the difficulty finding good models, up to this point, I've been using heavily merged combinations of three or four other models, to get the realistic look I wanted, though it wasn't quite perfect. Even with many layers of resampling, it had a very slight uncanny valley look.
Then, I found this model. Holy crap. The first image I created, without tweaking, using my existing multi-sampling workflow...my partner looked at it and their jaw dropped. No uncanny valley. Near-realism. I'm stunned.
Yes, one can get bad images from this, with a bad workflow. You can with any model. Occasionally, I have issues with certain attempts ignoring my prompts, but it's very specific ones and I think I've mapped out all my problem ones.
This model is stunning. I sincerely hope you continue to iterate on it!
If you have the processing power you could also take your images and put it through a Flux2 Klein 9B upscale pass to push the realism even further while keeping the furry anatomy. Which is something I've personally been playing around with a bit recently.
@willywizard Was actually working on that today! Great timing. :D
@willywizard Out of curiosity, what upscale factor, sampler, scheduler, and denoise did you end up with for the upscale pass? I keep thinking I found a good setup, but then the next image, klein all but destroys the image. Looking for something more or less consistent and figured I'd ask the master. :D
@LemurT You gotta use Klein in image edit mode if you are going to use it for upscale pass, otherwise it will destroy the image. Outside of that its up to personal preference and experimentation.
if it is in image edit mode and its still destroying your image, it likely has to do with your prompt or something else but I can't help you troubleshoot, just experiment with it.
I might publish some workflows soon that has examples but all the settings are close to default.
so whats different with core compaird to the other Willy's noob realism EPS and V-pred models? just wondering
Edit: nvm, i didn't read enough xD. so its built from the ground up. idk, i don't see too much of a diff in realism/quality from prevoius but it could just be me
Yeah, I just decided to separate it out. Probably shouldn't have, but because I built this model version from the ground up I didn't think it belonged as a different version of the older model which had different base recipe and was kind of inbred at that point.
Just trained a LoRA on this. Absolutely awesome! Almost trained itself!
What's the proper VAE to use with this?
The default SDXL vae
@willywizard Thank you 2x.
I see you updated the cover photo. Working on anything else? Been a while. Love your models.
I just swapped the two images in the showcase.
I am, but everything has been kind of slowing down or being more difficult than it really needs to be right now.
As much as I wanted to and tried to work on a better version of this for a while, I couldn't figure out any ways to further improve it right now so I am just either waiting for some newer/different stuff to come out to use as merging resource (if it even works), or when I have the time just really go deep into experimentation with it.
Most of what I am working with right now is using Flux 2 Klein 9B in image edit to upscale and enhance the images further, and other than that I have been playing around a bit with Anima and Krea 2.
Everyone has kind of seemingly branched out from SDXL at this point, and I completely understand why.