<h1 id="virile-stallion-pony-xl-lh77hs2yl">Virile Stallion - Pony XL</h1>Come be weird with us on <a target="_blank" rel="ugc" href="https://discord.gg/fmqp8WM4">Discord</a>! <h3 id="if-you'd-like-to-buy-me-a-beer-or-coffee!-wnx3tnvwq"><a target="_blank" rel="ugc" href="https://ko-fi.com/scratchproof">If you'd like to buy me a beer or coffee! 🍺☕</a></h3><h2 id="v5.0-photoreal-tss2u93dr">V5.0 - Photoreal</h2>Config I use:Sampling Method: DPM++ SDESampling Steps: 18 - 24CFG Scale: 4 - 6Hires Fix: OnUpscaler: 4x_NMKD-Superscale-SP_178000_GHires steps: 18Denoising strength: 0.25Upscale by: 1.8 (You don't have to go that high, just whatever looks best to you)Res: 768 x 1024Adetailer: Default Settings with face_yolov8n_v2.pt or face_yolov8n.pt<h2 id="v4.0-photoreal-experimental-lci9hestm">V4.0 - Photoreal - Experimental</h2>Config I use now: Sampling Method: DPM++ SDESampling Steps: 18 - 24CFG Scale: 4 - 6Hires Fix: OnUpscaler: 4x_NMKD-Superscale-SP_178000_GHires steps: 12Denoising strength: 0.3Upscale by: 1.8 (You don't have to go that high, just whatever looks best to you)Res: 768 x 1024Adetailer: Default Settings with face_yolov8n_v2.pt or face_yolov8n.ptA 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!<h2 id="v3.5-beta-cinematic-(scores-and-no-scores)-ns17o9llm">V3.5 - BETA - Cinematic (Scores and No Scores)</h2>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!<h2 id="v3-beta-cinematic-(scores-and-no-scores)-w4z75ohwq">V3 - BETA - Cinematic (Scores and No Scores)</h2>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!<h2 id="v3-beta-5e3wm92wm">V3 - BETA</h2><ul><li>Faces should be better</li><li>Should get a lot less watermarks</li><li>Realism should be more easily achieved</li></ul><h2 id="v2-beta-w33ans3bf">V2 - BETA</h2><h2 id="read-first-for-best-results-bhlsw7cu5">👇 Read First for best results 👇</h2><h3 id="preferred-settings-zc5r0hlxp">Preferred Settings</h3>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 <a target="_blank" rel="ugc" href="https://github.com/ahgsql/StyleSelectorXL">SDXL Styles extension</a> to pick the style I want. Works pretty well for me. You can also get an expansion for the extension with many more style <a target="_blank" rel="ugc" href="https://civarchive.com/models/132426/art-styles-expansion-for-styleselectorxl">here</a>.Clip Skip: 2 for Comfy (doesn't matter on A1111/Forge)VAE: sdxl_vae_fix.safetensorsSampler: DPM++ 2M SDE KarrasSampling Steps: 20+ (I use 38)Hires. Fix: OnUpscaler: 4x_NMKD-Superscale-SP_178000_GHires Steps: 16Denoising Strength: 0.35Upscale by: 1.5+Res: 1024x768CFG Scale: 4/5/6 (You can even try and push to 7 depending)Adetailer: Yes<h3 id="achieving-better-more-realistic-and-varied-faces-7xv5klf7q">Achieving better / more realistic and varied faces</h3>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 MeshAdetailer Settings:Adetailer Positive Prompt:<pre><code>face</code></pre><ul><li>User separate steps ✅<ul><li>28 steps</li></ul></li><li>Use separate CFG scale ✅<ul><li>0.35</li></ul></li><li>Use separate checkpoint ✅<ul><li><a target="_blank" rel="ugc" href="https://civarchive.com/models/337836/virile-xl">virileXL_alpha10.safetensors (or any other realistic model you like)</a></li><li>(Change Steps / CFG / Sampler accordingly if you use another model)</li></ul></li><li>Use separate sampler ✅<ul><li>DPM ++ 2M SDE Karras</li></ul></li></ul>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.<hr />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.<h2 id="general-settings-fstz0ogco">General Settings</h2>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:<pre><code>score_9, score_8_up, score_7_up [Your Prompt]</code></pre>OR<pre><code>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]</code></pre>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 TAGSI cannot recommend enough for you to install the <a target="_blank" rel="ugc" href="https://github.com/DominikDoom/a1111-sd-webui-tagcomplete.git">SD Webui Tag Complete extension</a> 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.<h2 id="gen-settings-uqumtvlnk">Gen Settings</h2>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 experimentVAE: sdxl_vae_fix.safetensorsSampler: DPM++ SDESampling Steps: 14Hires. Fix: OnUpscaler: 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: 14Denoising Strength: 0.5 with Latent / 0.25 with SuperscaleUpscale by: 2 with Latent / 1.75 with SuperscaleRes: 1024x768CFG Scale: 3/4Adetailer: YesAdetailer Model: mediapipe_face_mesh or Yolov (Mediapipe looks more natural but fails easily on faces far away)Adetailer Settings: DefaultThat 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.

Virile Stallion

V4.0 Photoreal<ul><li>Fairly experimental</li><li>Boosted the realism</li><li>Lot less "same face" issues</li><li>Prompt adherence might be slightly reduced</li><li>Expect some funny stuff</li><li>Uses a mix of natural prompting and booru tags</li></ul>

Files with SHA256: 049023be57c1bafa84e83a5009da25eeb92511eca779515b72544de9d949e84c