Virile Stallion - Pony XL
Come be weird with us on Discord!
If you'd like to buy me a beer or coffee! 🍺☕
V5.0 - Photoreal
Config I use:
Sampling Method: DPM++ SDE
Sampling Steps: 18 - 24
CFG Scale: 4 - 6
Hires Fix: On
Upscaler: 4x_NMKD-Superscale-SP_178000_G
Hires steps: 18
Denoising strength: 0.25
Upscale by: 1.8 (You don't have to go that high, just whatever looks best to you)
Res: 768 x 1024
Adetailer: Default Settings with face_yolov8n_v2.pt or face_yolov8n.pt
V4.0 - Photoreal - Experimental
Config I use now:
Sampling Method: DPM++ SDE
Sampling Steps: 18 - 24
CFG Scale: 4 - 6
Hires Fix: On
Upscaler: 4x_NMKD-Superscale-SP_178000_G
Hires steps: 12
Denoising strength: 0.3
Upscale by: 1.8 (You don't have to go that high, just whatever looks best to you)
Res: 768 x 1024
Adetailer: Default Settings with face_yolov8n_v2.pt or face_yolov8n.pt
A reworked merged focused on realism - This comes at the expense of some prompt adherence but you can still try and experiment.
Except some funny looking stuff at times.
Faces can sometimes look a bit funky, especially when more than 1 subject in the picture. I'll try and rework that in the future.
Negative prompt is important. It might make or break your gen when it comes to realism. Check out the ones I used and work from there.
I use low config and the Euler sampler (Not Euler A).
You can also use DPM++ SDE which should give much better compos.
Just see what works for you.
Enjoy!
V3.5 - BETA - Cinematic (Scores and No Scores)
More training, reworked faces and slightly altered style.
Works fairly nicely without Pony scores (just use "realistic" in your prompt).
Negative prompt also has quite a bit of an impact, so play around with it.
Have fun!
V3 - BETA - Cinematic (Scores and No Scores)
A different style that I personally like.
Also works very nicely without the need for Pony scores (score_9_up, blabla, etc.).
Just experiment and have fun!
V3 - BETA
Faces should be better
Should get a lot less watermarks
Realism should be more easily achieved
V2 - BETA
👇 Read First for best results 👇
Preferred Settings
NB: I find that I am getting better results without the usual Pony negatives.
I like to only set up my Pony Positives (score_9, score_8_up,score_7_up, score_6_up, score_5_up, score_4_up), then my prompt, leave the negatives empty and then use the SDXL Styles extension to pick the style I want. Works pretty well for me. You can also get an expansion for the extension with many more style here.
Clip Skip: 2 for Comfy (doesn't matter on A1111/Forge)
VAE: sdxl_vae_fix.safetensors
Sampler: DPM++ 2M SDE Karras
Sampling Steps: 20+ (I use 38)
Hires. Fix: On
Upscaler: 4x_NMKD-Superscale-SP_178000_G
Hires Steps: 16
Denoising Strength: 0.35
Upscale by: 1.5+
Res: 1024x768
CFG Scale: 4/5/6 (You can even try and push to 7 depending)
Adetailer: Yes
Achieving better / more realistic and varied faces
Default faces have a tendency to all look the same and are not the most realistic (this is due to Pony).
The best way to circumvent this is to use Adetailer and to configure it to use another model in order to inpaint the faces.
Adetailer Model: face_yolov8n_v2.pt OR Mediapipe Mesh
Adetailer Settings:
Adetailer Positive Prompt:
faceUser separate steps ✅
28 steps
Use separate CFG scale ✅
0.35
Use separate checkpoint ✅
virileXL_alpha10.safetensors (or any other realistic model you like)
(Change Steps / CFG / Sampler accordingly if you use another model)
Use separate sampler ✅
DPM ++ 2M SDE Karras
Other settings can remain the same. This will allow Adetailer to switch settings and model before restoring the faces and will ensure that you get much nicer, varied, realistic faces.
Virile Stallion is my first Pony merge. It attempts to offer realistic outputs while keeping Pony's flexibility when it comes to anatomy, concepts and prompting.
This is an alpha release, meaning that I am fully aware that improvements can be made. I intend to keep on working on this model as I go, to bring as many improvements as I can.
This model is versatile and will work for both male and female characters.
It was build around Virile XL as well as Pony XL.
General Settings
Model should work without the usual Pony Strings such as "up_this" and "up_that", HOWEVER, i find that outputs are better and more detailed with the following in the positive prompts:
score_9, score_8_up, score_7_up [Your Prompt]OR
Positive: (score_9, score_8_up:1.1), score_7_up, realistic picture, [Your Prompt]
Negative: score_6, score_5, score_4, [Whatever you want]The aim was for the model to work out the box without the need for finicky prompts. You can however play around with those if you wish as they should still affect the outputs. You may even find better ones.
THIS MODEL WILL WORK BEST WITH BOORU TAGS
I cannot recommend enough for you to install the SD Webui Tag Complete extension for A1111 or Forge to make your life much easier.
I have found that some Loras do work with the model. They are however a bit finicky and may need a high CFG to work. Just try and see what you get.
Gen Settings
You can find the settings I use to gen by looking at the showcase pictures, and here below, if you want similar outputs:
Clip Skip: 1 (I believe it looks better) but experiment
VAE: sdxl_vae_fix.safetensors
Sampler: DPM++ SDE
Sampling Steps: 14
Hires. Fix: On
Upscaler: Latent (Nearest-Exact) [Will be smoother with less details and a tad less realistic] OR 4x_NMKD-Superscale-SP_178000_G [Slower, but more realistic outputs]
Hires Steps: 14
Denoising Strength: 0.5 with Latent / 0.25 with Superscale
Upscale by: 2 with Latent / 1.75 with Superscale
Res: 1024x768
CFG Scale: 3/4
Adetailer: Yes
Adetailer Model: mediapipe_face_mesh or Yolov (Mediapipe looks more natural but fails easily on faces far away)
Adetailer Settings: Default
That should be it. Then check out the Showcase for Pos and Neg prompts.
Again, this is an alpha model, it is not perfect, expect some weird stuff.
In any case, I hope you enjoy it and I look forward to seeing what you make of it.
Description
V4.0 Photoreal
Fairly experimental
Boosted the realism
Lot less "same face" issues
Prompt adherence might be slightly reduced
Expect some funny stuff
Uses a mix of natural prompting and booru tags
FAQ
Comments (34)
The results are nice and I have a lot of use cases for this. Thank you!
Dont forget Virile XL Alpha 2!
First off - thank u for ur labor 💪 this model looks great n having pony-powers is awesome!
Unfortunately, I haven't been able to replicate results in A1111 n I can't use comfy
Am I missing something? Cuz even with the exact same input my output is significantly different.
hi... how can i achieve furries with this model please? :')
Hi. So far great model. Thanks.
As a heads up and just for info.
Using words and phrases that Pony or the checkpoint doesn’t know causes a few issues, the main being Deformity _ especially the 3 finger problem. That's not bashing the model it’s a diffusion issue.
Pony tends to make the deformity issue more obvious, deleting unknown words tends to fix the issue.
The second huge issue is using sexual terms that pony has been trained to recognise as hetero sexual acts.
If it’s a common term or one that pony doesn’t know it defaults to using the female form which is why you get transgender / futanari constantly.
Any common sexual term will also try to default to a female.
Anal Sex - Pony knows Anal and it knows Sex but not together, pony sees Sex and tries to default to hetero sex, so you get a female head on male body.
Pony is dominant in the checkpoint so is over powering the Anal Sex tagging which leads to a lot of vaginal male insertion.
Gay, Homosexual, Homo are not well trained and tend to add to deformity.
Woman and Man or Male and Female have always been a diffusion issue, diffusion defaults to the female as it’s over trained on everything feminine.
Which is why Pony uses 1boy, Bara and Yaoi. The words aren’t common western hetero terms for Woman, female etc, and haven’t been nicked yet like all the other gay words and doings lol.
1man – Pony tends to default to woman without a fine tuned lora using the 1man term.
Hopefully someone finds that useful
You have surpassed yourself. I think you did it really well. I think the results are great and awesome. Thank you, thank you, thank you. I love you for that :)
I don't know how, but this model (version 2.0) is the best pony model I've seen so far. Amazing! Thank you for it <3
This model fuckin rocks 💪
V2 observations:
- adjusting ADetailer like described in the instructions really improves faces and cocks, especially when using an SDXL LoRA in the ADetailer prompt. I also found RealisticVisionXL (specifically realvisXLlightning) to be great at this, as long as you adjust the settings correctly
- running the refiner w VirileXL seems to make a positive difference. I did 0.8 n it worked well, as long as there were no pony LoRA in the prompt to complicate things. I know it's not the best way but I'm allergic to upscaling 🤷
- I used character LoRA for pony to render then character LoRA for SDXL to ADetailer n the results were 🔥🔥🔥🔥
Quick note: Having anything from score_1 to score_3 in negatives does not do anything as the pony model was trained down to 4 only. Unless you added some really bad images in the dataset and custom tagged them which I'm not aware of.
This looks amazing!!! Can you upload to your Tensor.Art page?
Hi! this model is amazing, but my gens are not realistic enough, the pics are somehow more like game cg, rather than photorealistic ones, any advice? trying to use the detail tweaker lora, the effect is limited, anyway thanks for the great model!
fuck this model is so good. 11/10
V3: much better faces for real. Something I've noticed is including "explicit"and/or "rating explicit" and/or "uncensored" in the prompt with "censored" in the negatives really improves the dicks. Seems more relevant w V3 than V1 n V2. Gets rid of lil artifacts that, I assume, try to be censor bars cuz Pony.
Wow this looks really good! Wish i can try it haha
Best model for real, even though the clothing options like rubber suit or something like that is not available.
This is the only Pony model where men's faces look very good and where there is good photorealism!
Nice improvement on 3.5! Much easier to get different faces, no need to use ADetailer.
Really awesome work! Hope to see virile fusion for pony too!
Would it be possible to attempt a hyper or lightning version?
You model in my opinion is the best out there! Always check to see when there's an update. 4.0 seems the best in terms of quality so far but you REALLY gotta push it to get anything NSFW. Other than that wow, amazing work!
This is a really good model, I hope you keep improving upon it.
As always, wonderful model, and thanks for the frequent update! Renders perfect male form!
Just a notice, that the up-side-down human face is not well drawn. I guess it's a common problem for basically all models, but is there a trick or sth to make it better?
I love how much work you're still putting into your models, Scratch. I wouldn't have tried SDXL and PONY had you not worked on these first. Thanks so much for all the great stuff 😎
Thanks all of your Work from SD1.5, SDXL and Pony. It's AMAZING!
Absolutely amazing checkpoint - Enough pony prompt adherence to be precise yet not so much pony that it makes it look unrealistic. Fantastic job!
Man I swear I felt this update in the force or something. Frothing with excitement - thanks for all your labor 💪
Hello good job.I have a question:
What is the main difference between Valliant Stallion and Virille Stallion? And the ability to create women/trans content? thx
5.0 seems superior in every way than 4.0. Haven't been able to use that much but so far it looks better, understands better, and even NSFW comes out easier. Always amazed by your work!
feel like this is pushing pony too far. seeing lots of floating penises and bad anatomy in v5.
Amazing, indispensable model. I do wish it had better prompt adherence wrt different ethnicities and body types. I have a lot of generations where the prompt is like "slim Korean man" or "gangly Persian man" and it generates a blond Abercrombie model.
work good with dmd2 lora, lcm, step5,cfg1.3
测试下来最强的pony，而且对controlnet的效果超过的其他所有pony,XL模型。甚至超过了老版1.5的效果。你的模型真正开启了全面超越，太感谢了。
one of the best set around. Thanks
game changer - thank you!
Hello ! Quick question. Why recomment using the DPM++ instead of the DPM++ 2M or more recent ? Is there something which works less well with more recent dpm samplers ?