This model represents the original Willy’s Noob Realism line and is now considered the Legacy branch, while ongoing development has moved to Willy's Realism.
This model and its version(s) were merged for the sole purpose of finally having a decent out of the box consistent Realism model based on Noob. The "HD" and "V-Pred" versions are made for higher resolution raw gens and upscaling.
!!! If you are having problems, noisy images, or prompt issues please scroll down to "Common Problems" !!!
Version Information
V-Pred v2.0
This is the second version of Willy's Noob Realism V-Pred.
This version is almost a hybrid between the old style of model and the newer V-Pred v1.0, allowing for both darker scenes, and bright natural lighting scenes similar to that of the old EPS models and the V-Pred alpha.
As I slowly continue to try to improve on my models, I've found it getting harder and harder to improve the quality of the model without causing significant other issues or significant notable loss of knowledge in certain areas.
In this version I've replaced the clip-L and tweaked the models using my own and other specific loras.
This version works well with CFG 4-9. The lower the CFG the more balanced the colors, working better with darker scenes, and higher CFG works better with saturated colors and bright scenes.
V-Pred v1.0
This is the first release version of Willy’s Noob Realism V-Pred.
This version is a bit different from the earlier versions of Willy's Noob Realism, so you can expect it to behave differently in some ways. While there’s a bit of style loss compared to before, this version keeps the high resolution detail and realism—and, most importantly, it reintroduces proper lighting overall, which every previous version struggled with.
Getting this V-Pred model where I want it has been way tougher than I expected. What I thought would be a quick process turned into weeks of experimenting just to find a way to bring proper lighting back into a model that seemed almost incapable of handling it. (For perspective, getting from v4 to V-Pred alpha only took a single day—everything after that has been the real grind.)
What I’ve learned along the way is that both the style and the lighting issues both sit in the same weight layers. That makes fixing them tricky, since the options are either doing some pretty janky merging, or spending a ton of time diving deeper into each layer and many hours of experimentation.
This version works well with a CFG of 5-7. Lower CFG is better for darker scenes while higher CFG is better for brighter scenes and more saturated colors.
V-Pred Alpha
V-Pred Alpha is the first version of Willy's Noob Realism that uses v-pred instead of eps noise prediction.
This version is capable of more vibrant colors, and overall darker, and lighter images.
Being the Alpha version, its far from perfect. Fixing the "spotlight" issue from eps, and training back in proper dark lighting for dark scenes was taking significantly longer than I anticipated, so I decided to release the initial version now rather than making people wait longer while I continue to play around with datasets and training settings.
V-Pred models don't typically play nice with play nice with a lot of sampler & scheduler combinations. So with this version you might have to play around to find what works right, or you can go back to v1.0's recommended sampling settings which work perfectly fine.
HD v4.0
HD v4.0 is the newest and final version of the "HD" series of Willy's Noob Realism models. This version was made as a final attempt to increase realism, while reinforcing previously lost knowledge and further stabilizing the model.
This version can generate realism without the usual required quality and style tags, and does not need any negative prompts.
Where do we go from here?: My next goal is to make a proper v-prediction version of Willy's Noob Realism.
HD v3.0
This version was merged to further increase realism, allowing for more variety and expressiveness. Fixed foliage and most backgrounds while limiting excess noise.
This version has been tested with both Euler A & Beta, and DPM++ 2m SDE & Karras and works with both tag based prompting, and some natural language prompting.
(Works best with just tags, or a combination of tags and natural language)
HD v3.0 has improved stability at higher resolutions, capable of fully stable 1920x1920 raw generations, and 2048x2048 raw generations with minimal to moderate issues depending on prompt and aspect ratio. Raising the CFG can increase stability further with problematic prompts at 2048x2048. (Skimmed CFG can be used to decrease oversaturation caused by high CFG)
HD v2.0
This version was merged to increase realism, while attempting to fix the problems with noisy backgrounds, while retaining its knowledge of complex topics and species. Can generate raw images up to 1920x1920 with minimal issues.
Uses the same version specific settings as HD v1.
HD v1.0
This version was made to increase realism, while allowing for higher resolution base gens with the quality of an upscaled image! This version is capable of generating images from 768x768 to 1920x1920 resolutions without breaking.
Version Specific Settings:
Prompting is the same as the other models, but to get the most detail out of higher resolution images generating images with at least 50 steps is suggested. Please refer to the "Recommended Settings" for more information.
If you are getting images with unintended multiple characters, make sure to include the tag "solo".
Longer, more detailed prompts result in more stable and typically more detailed images.
Certain prompts may not be fully stable in this version, so play around.
v2.0
This version uses the same settings as v1.0, and is an overall improvement in realism, and better support for upscaling.
v1.0
Initial release version.
Recommended Settings:
TL:DR Its very simple, just prompt however you usually would for noob, but add "realistic", and "photorealism" after your quality tags.
I suggest using Euler A with Beta scheduler, or DPM++ 2m SDE with Karras scheduler (does not work with v-pred) at a CFG of 5-7 and steps of 20-50, but feel free to experiment. Not all schedulers and settings work for everyone.
Here's a positive prompt example:
<quality tags>, realistic, photorealistic, <style tags>, <character tags>, <pose tags>, <background tags> And here's a basic filled in example:
masterpiece, best quality, highres, absurdres, (realistic, photorealistic), 1girl, solo, anthro, wolf, grey fur, posing, bedroomAnd here is my suggested negative prompt:
(Negative prompt is no longer needed for quality images)
worst quality, low quality, normal quality, monochrome, bad anatomy, ugly, (illustration, cartoon, anime), flat color, outline, smooth skin, blurry,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 aren't set in stone, just my own personal preferences from testing and using the model in ComfyUI
Unless you specify a gender, it will automatically generate female characters or characters with breasts due to its bias towards women.
The tag "1girl" can cause issues with certain prompts when trying to generate realistic humans. If you are getting strange outputs, try just using the tag "female" instead.
If you are getting unwanted multiple characters, make sure to include the tag "solo"
Beta scheduler is not properly implemented in Forge and will result in heavily artifacted noisy images.
Here is a list of working Schedulers for Forge: https://civarchive.com/posts/11756794
From my testing Beta scheduler does work in Reforge however some have been experiencing problems. I suggest you make sure Reforge is up to date before complaining to me. Otherwise most other samplers work perfectly fine in Reforge.
Here is a list of working Schedulers for Reforge: https://civarchive.com/posts/11756585
Nobody using A1111 has actually expressed any issues. If you are having issues with A1111, I would suggest verifying your settings.
Here is a list of working Schedulers for A1111: https://civarchive.com/posts/11757173
Some people have also been experiencing small VAE artifacts while using Reforge. This is likely due to your VAE settings not being set correctly. This model uses the default sdxl.vae.safetensors sdxl VAE.
Some EPS loras can cause problems with the V-PRED versions of this model such as the EPS "spotlight" effect, making night/dark lighting brighter, or colors less saturated.
Other information:
This model was merged with furries in mind, but is fully capable of generating all kinds of realistic humans and humanoid fantasy species. I have extensively tested it with all kinds of different unique furry species, all kinds of humanoid fantasy creatures, and humans. I have also tested its compatibility with SDXL, PDXL, IL, Noob loras. Most loras trained on proper tags will work with the model, although Noob loras work best (since that is the base model for this merge), and other SDXL based model loras will likely work with varying success.
"Fake slider" Pony loras will not work, as those are trained on Pony's quality tags for activation (There are very few of these, so just be warned).
I plan on making better versions of this over time as more resources come out, or if I have time to make some realism or other Loras to further fortify it.
If you have any issues, feedback, or suggestions feel free to let me know in the comments!
Description
This is the first release version of Willy’s Noob Realism V-Pred.
The V-Pred alpha was essentially just a straight conversion from v4, but this release takes things further. It’s quite a bit different from the earlier versions of Willy's Noob Realism, so you can expect it to behave differently in some ways. While there’s a bit of style loss compared to before, this version keeps the high resolution detail and realism—and, most importantly, it reintroduces proper lighting overall, which every previous version struggled with.
FAQ
Comments (22)
when will they be available/out of bid again?
The review for release version of vPred v1:
I still had been using v1 (the very first version) and mostly will compare it with v1. vPred have more saturated colors but unlike all other versions, it's not just oversaturated, it actually look beautiful and some images are really stunning! Also, the background, clothes, character and the image overall seem to have much more details, especially background. But it's quite bad at diverse feral species, every non-specified species are treated as anthro, and feral species anatomy (body shape) is often some way wrong (it looks like some "average feral body shape" and some species like bird, kangaroo, probably also any bipedal feral, snake or other that have "not average mammal" anatomy may look wrong, especially at certain angles and tags).
Try this one:
prompt:
bird, {feral, |}raven, bipedal, rear view, from below, depth of field, epic, dark fantasy,negative:
(human), anthro, (tits, breasts:0.6), no tail, tailless, muscular male, guro, ero guro, cigarette, high contrast, quadropedal, mammal, nsfw, butt, hands,First, you notice that a character is never a feral bipedal bird – it's either anthro or quadropedal (when feral is present.)
Second, it''s always nsfw. Adding anus, pussy, penis, balls to negative helps a bit, but still a lot of nsfw results.
Well, and third is a bit too extreme black and contrast, but it will go if remove dark fantasy tag, the colors without this tag are fine.
If you'll try v1 for the same prompt, you'll get bipedal sfw feral raven every time when feral is present (also, no need of bipedal positive and anthro, quadropedal, mammal, butt, hands negative, but these (except hands) didn't helped vPred anyway) and sfw anthro raven when not (and no extreme black levels).
I may be a bit wrong about something, because I generated just a little amount of images with it (200-300) and compared even less with v1 (to make sure it's a model version specific, not bad prompting or anything else)
Yeah. i know this version isn't perfect. Always making everything NSFW which is why i didnt put a whole lot of SFW in the showcase.
However, the main issue is that for some reason gender tags cause NSFW, if you remove gender tags it stops constantly putting genitals on everything.
The only good furry realism model AND only good Vpred model. Nice combo br0h
Going to be testing 1.0 vs alpha, but what is the major difference between the two?
alpha is basically a straight across conversion. The closest I could keep it to hd v4 without nuking anything while making sure it was working while using vpred ztsnr.
with v1, I basically took alpha, and tried to add back in the knowledge of proper dark, varied types of lighting which ended up being more difficult than I originally thought since the style info and some other important knowledge was effected, so I also tried to re-balance it which changed the overall style a bit.
since alpha was basically just hd v4 converted it still had the "spotlight" issue with dark/night scenes and struggled with proper lighting, and v1 was several iterations into me trying to fix that.
Doesn't seem to be working in A1111 or Forge. Just generates a very noisy image. Possibly needs a .yaml file? I'm not too familiar with V-Pred models
Edit: Forge works fine, just can't use 'Beta' schedule
A1111 is pretty outdated and missing a lot of features which is why most people moved to Forge, Reforge, and ComfyUI.
I personally don't use Forge because their implementation of the Beta scheduler was broken for a long time, but I was told that they recently made changes to fix it.
However the model should work perfectly fine using the recommended settings in both Reforge, and ComfyUI.
Did tensor hide your checkpoints or do they need different cover art?
No, is still there. I just haven’t updated anything over there since HD v3, and you need to be logged in to see it.
@willywizard Yeah, I'm logged in. it shows you have 3 models & 3 posts but I see nothing. edit: my account is before the 18+ toggle was removed.
All I get is noise. Followed prompt and setting guidelines posted with model.
what UI are you using? Base A111 doesnt support vpred and just outputs noise, if thats what you are using.
it's a Vpred model, unless you have tweaked your inference UI accordingly, it will only generate grains and noise...
[DO AT YOUR OWN RISK]
[Some forks of A1111 are not compatible with Vpred models for various reasons.]
A1111 relies on YAML files for detecting if a model is EPS or Vpred, they should be present as such:
models/
└── Stable-diffusion/
├── example1.safetensors
├── example1.yaml ← must be here
├── example2.safetensors
└── example2.yaml ← must be here
add these to your launch parameters:
--no-half-vae
"Almost all SD 2.x and many SDXL v-pred models + their VAEs will silently force epsilon mode or produce black images in half-precision. This flag fixes it instantly."
--disable-nan-check
"v-prediction naturally creates higher-magnitude latents → the old detector kills generations with false-positive NaN errors (especially CFG > 8 and Euler/DPM++ SDE samplers)."
--v-parameterization
"Safety net for when a model author forgot to include the correct .yaml, renamed it wrong, or you’re using a merged model that lost its config. Overrides everything and forces v-pred mode."
Really weird issue on comfyui:
When it hits the Ultimate SD upscale nodes, it creates a whole different image and doesn't even seem to upscale the original image at all
That is an issue with your settings or the node, not my model. I'm not general tech support.
@willywizard it only seems to do it with your model so I thought I'd bring it up, wasn't tryina be a cunt
how good is this model at generating text/dialogue?
just like most SDXL models, it’s hit or miss. Can sometimes do short texts pretty good, but can’t do much than that. Single words it can usually do fine, or very popular groups of words. Like “happy birthday”
what settings do you recommend for img2img?
Honestly I don't do a whole lot of img2img, but the settings will depend on what you are doing with it.
Usually if you are trying to do something like an img2img style transfer (anime image -> realistic image) you would want to use a denoise of about 50% or a bit lower. If it starts to change things you don't want it to like anatomy, you can stack a canny or line art control net on top of it to a with an end point of 60-70%.
If you aren't trying to go for anime to realism style transfer and just trying to upscale or improve another realistic image using my model. Something like 35-40% denoise will probably do the trick.
For the rest of the generation settings, the default suggested ones or any other combination that works to generate normally with my model should work fine.
@willywizard thanks for the response