Bilitis' Garden of Pleasure
Release Date: July 7 2026
https://milovana.com/webteases/showtease.php?id=82052
Venture into my garden where your mind will be flooded with erotic hallucinations. The only way out is to prove your cognitive abilities in 5 different trials.
Features
- 4 main endings (and maybe a secret one...)
- Randomization to make it enjoyable multiple times
Muses
Technical Details and Credits
For other creators or those who are curious.
- Spoiler: show
-
Visuals
Visual assets (not photos) were generated with Gemini, sometimes with some additional manual editing. Gemini seems a bit inconsistent with it's NSFW filter. It will generally refuse to generate anything too explicit, but sometimes it surprises you.
For the "Sexy Spins" trial, I generated a few different circular assets with Gemini and then wrote a Python Colab to generate a bunch of rotations of each with random lighting added to make it harder to rely on shadows and highlights to beat the game.
Photos
All photos were sourced from Elite Babes. I started off manually downloading photos I liked, but eventually wrote a Python script to crawl different models, studios, and tags for gallery urls and then download the photos in bulk. This was especially useful for the "tease" segments where I just needed lots of randomized photos for variation. For those I wrote another script to sort them into folders based on content using Llama JoyCaption, which is a fantastic image classification model that can handle NSFW content. This quickly gave me 1000+ photos sorted into folders to put into the tease. Llama JoyCaption was also great for filtering. For example, in this particular tease I mostly wanted solo scenes and no men, which the model was easily able to check for. You can run the model in Colab with the free GPU runtime, which is super convenient.
Music
The music is AI generated and royalty free. The Bilitis intro track was generated with Gemini by me. The other tracks are from AITunez Music.
Voice Lines
The voice lines were all recorded by me and then modified with Applio. The reasons for editing the recordings was to maintain my anonymity and to make them more ASMR-like (I unfortunately am not a professional voice actress). I first played around with text-to-voice models, but most of them are not well-tuned for erotic content or will even block it altogether. With erotic audio, I think cadence is just as important, maybe even more important than the words themselves—no cocks are getting hard nor pussies getting wet from a podcast voice or something straight out of a TED Talk, no matter how dirty the lines are. I found that even with careful prompting, mainstream text-to-voice models do not seem to be able to reliable generate sensual whispers.
With voice-to-voice models like Applio, you have much better control than with text-to-speech. The only downside is that you actually have to record all the lines, but I found that part quite fun. I set up a Python Colab to process files in bulk, so the conversion was very painless and the quality was mostly good (I tossed out a few files, but most were fine). I found a pre-trained model that was kind of similar to my own so I was able to get good results without any pitch shifting, but Applio should be able to do that too.
I was originally planning to add voice lines for all dialogue, but that felt like too much work, at least for this time. Also, I felt like I was pushing the limits of the EoS engine already with all the assets.
I figured out some ways to improve the quality later during development, but I was too lazy to redo all the lines. Expect cleaner voice line in my next tease 
Scripting
Using EoS for a project that relies on so much randomization was a fun challenge. My approach to avoid thousands of if-statements everywhere was to keep the state in the javascript as much as possible, and rely on "hub" pages that redirect based on javascript functions.
For example, a "game hub" checks the code for the current trial and round number. There's a "tease hub" that selects a folder to show a random image from and randomly decides if a voice line should be played or not. There's a "voice hub" that plays a random voice line that fits the selected image folder, and then immediately redirects back to the "tease hub" that actually shows the image. After the image timer runs out, you're redirected back to the "game hub" where the next game state is shown.
This approach turned out to be easy to work with and scalable. Although I wonder if the redirecting might make the tease run a little slowly if done excessively.
A tricky part with this approach is that you sometimes need to make a lot of nearly identical pages since, as far as I know, there is no way to directly pick what image to show or what audio file to play with the javascript code; you either have to have a bunch of if-statements or a bunch of pages. In two particular cases: the rotating objects and the voice lines, I actually downloaded the json of the project and added pages to that directly with a Python script rather than editing in the EoS UI to speed up the process.
There's probably lots more about the coding I can share, but I'll stop here. I'm happy to answer questions and also please feel free to open up the project directly if you want to see how something was implemented.
Special Thanks
for beta testing the tease.
Please leave a comment in the thread if you're interested in beta testing future teases from me. It's not going to be anything particularly organized, so no promises, but I might send out a few DMs whenever I have something ready.
Other Notes from the Creator
After making
Whoredle, I had two goals in mind for this tease:
- Make the game elements less time consuming
I found a big weakness with Whoredle to be that the game element requires too much focus and takes up too much time, making it difficult to keep the player aroused. I also received feedback that word games like that can be quiet inaccessible, or at least unfairly difficult, if English is not your first language. This time I wanted to make something more accessible.
I liked the idea of cognitive tasks, mostly involving memory and pattern recognition, as a way to flood the player's mind while teasing them. All trial rounds are intentionally very quick and losing is not meant to feel overly punishing (you just try again with the next game).
However, a lot of the games actually turned out significantly more difficult than I had originally expected. Especially the "Color Climax" game is pretty crazy—during testing I don't think I made it through the entire Hard difficulty version a single time, even after reducing the difficulty slightly from my first implementation. I am curious to hear how others experience this. Is it too overwhelming to stay aroused or is the cognitive load combined with sexual arousal enjoyable?
Either way, for my next project I want to experiment with game mechanics that are less demanding.
I experimented a lot with AI generated assets in this tease, which I know some people do not like. I am not a huge fan of "AI art" myself, so I completely understand. However, for a hobby project like this I thought it would bring benefits that I could not feasibly get any other way (if I had the budget, I would obviously rather hire artists and voice actors). In my next project, I plan to rely less on AI generated assets, but might use it for visual. I personally draw the line at AI generated porn, which is something I am not interested in, but for logos, non-porn visuals, etc. I am fine with it. Curious to hear what people think.
