Polo - Suddenly Gay

by GymDreams
Two men, two horses, and an AI generation nightmare with 12 legs. Rendered in Virile Animation.
Polo - Suddenly Gay

A few men got distracted during a game of polo when they saw how attractive their opponents were.

Thematically related to Rugby: Suddenly Gay, this image set features virile men playing a game of polo, but got distracted by the other men during the game. They proceed to steal kisses while the horses are still running, climb onto other horses, and spend time away from the game altogether.

This series is intended to be playful and somewhat funny. Anything remotely gay is the last thing you’d expect from polo! The concept of this series came from a conversation with a dear friend. He doesn’t want to be identified, but said that I could credit him as “BB Love” — this name is our little secret, but so that will be! BB Love is a wonderful muse, and I get inspired everyday to explore new ideas and concepts I have not considered before. Thank you!

Visually, it’s the first time I have tried using the anime style — made possible by Virile Animation (Civitai), a new Stable Diffusion checkpoint model from the author Scratchproof, who is the brainchild behind Virile Fusion, Virile Reality, and Virile Motion. Virile animation is Scratchproof’s foray into 2D drawings and anime. Like all his other models, it was built with a strong male focus that is suitable for making gay AI works. As a bonus, it often renders men with very funny facial expressions that I have grown to love.


Before starting the renderings, I already had a hunch that polo could be a challenging subject matter to create. The engines can render horses fine, but the amin issue is the number of limbs involved in the same image.

Two men + two horses = twelve legs of nightmare

As I have written before, AI generation tech — specifically, diffusion txt2img tech — does not render images based on bone strcuture liked 3D modeling. In turn, they have a really hard time coming up with compositions that are logical.

There isn’t a good way to get around this issue unless you use Control Net, but the downside of using Control Nets to control compositions is that you’ll be constraining its compositional creativity. So I went with the more time-consuming route — create workable text prompts but generate a ton of images so that I can weed out the problematic ones afterwards.

The funny thing about this exercise is that the limbs were the least of my problems. I rendered at 768x768 with hires fix at 2x, but it’s still quite a tight space with so many figures, so many images actually turned into Centaurs — with human head and horse legs. Some people have once asked me how they could render Centaurs, because they could never do it successfully. Well, it would appear that prompting with enough men and horses and you will get them even trying!

I shared one of them in these sets, but if I have time I will include more failed renders in the future.

Since I made this set mostly to test the checkpoint model, the settings I used largely resembles the recommended settings on the Virile Animation page on Civitai: DPM++ SDE Karras, 30 steps, 768x768, 5 CFG cale. Hires upscale 2x, 15 steps, R-ESRGAN 4x+ Anime6B, Denoising 0.5. Clip skip: 2. ADetailer model: Post: Lightroom color correction, Gigapixel HQ 2x. If you prefer a more 2D look, the Euler a sampler also works very well for this model.

Text prompts in Stable Diffusion.

Set 1

Set 2

Set 3

Techncial Details

  • Steps: 30
  • Sampler: DPM++ SDE Karras
  • CFG scale: 5
  • Size: 768x768
  • Model hash: 59bcb95ef6
  • Model: virileAnimation_v10
  • Denoising strength: 0.5
  • Clip skip: 2
  • ADetailer model:
  • ADetailer confidence: 0.3
  • ADetailer dilate/erode: 4
  • ADetailer mask blur: 4
  • ADetailer denoising strength: 0.4
  • ADetailer inpaint only masked: True
  • ADetailer inpaint padding: 32
  • ADetailer version: 23.7.1
  • Hires upscale: 2
  • Hires steps: 15
  • Hires upscaler: R-ESRGAN 4x+ Anime6B
  • Version: v1.3.