Ren'Py Lab Rats 2 - Reformulate [v2025 4.0.37]

[Deane9850]

Will you be explaining why, or are we supposed to guess?

The current appointment system is currently broken.

 

[bowobble]

Also, model switching from the GenAI mod config screen might not currently work (I commented out some stuff that was giving me issues, but didn't have time to debug) - I use the a1111 Web GUI to switch the model, and just ignore the model in the mod config for now.

That's a good idea. I like to keep my models sorted, so I put them in folders but GenAI either escapes or doesn't escape the slash, so stable diffusion doesn't recognize the path to the model. Makes switching models easier for me.

Just a tip, especially if you create a lot of images: Change settings in webui to save images as jpeg instead of png (WebUI: Settings -> Search -> File format for images), my GenAI folder is 400MB big and that's from banging like 5 chars.
I'm using different triggers, like more positions (anal piledriver, skull fuck etc.), cum stains, or changing the background to the current location. This causes an exponential increase in generated images. Not sure how many images your modification generate.

 

[SMC85]

how do i open these in a way thats editable?
when i open with notepad the text just shows up as a bunch of "
¿ÿÑø;¯ÌÜÿí—Ë?àO²<e@?¸ ×wðuži±š" type symbols

There are two types of files baring the same name, .rpyc and . py files. .py files can be edited and then as the game starts will be transformed into the .rpyc files.

Edit:

I just read that there's a GenAI mod for different models? Where can I get that?

 

[bowobble]

I just read that there's a GenAI mod for different models? Where can I get that?

There is a guide to install Stable Diffusion, and here is the mod. You need to run the model locally, if you don't want that you need to change the mod and add the endpoint+credentials yourself.

 

[themagiman]

I have been playing with the GenAI mod posted above, and tried to modify it to use img2img instead of text2img generation based on the original images in the game, which turned out to be relatively simple to do, and yielded pretty decent results - it's not perfect and required some prompt tweaking, but overall the quality of the graphics is improved IMO, while keeping a lot of the charm and character of the original. Below are some examples of original images, and what StableDiffusion was able to generate from them:

View attachment 4812057 View attachment 4812058 View attachment 4812059 View attachment 4812060 View attachment 4812063 View attachment 4812064 View attachment 4812065 View attachment 4812066 View attachment 4812110

I feed character name into the prompt, and its surprisingly consistent in the face it generates for each character, while also being able to generate diversity of types and faces. Using img2img also more closely follows the outfits (and poses as well) in the game, where the text2img tended to generate its own outfits based on the prompt, which were not always very similar, and kept changing with poses. Each approach of course has its advantages, well setup text2img can also be fun to play.

Also I have been using SDXL Lightning models to do this, which can generate usable results in as little as 4 steps, much faster than regular models. So far my best results were obtained with

You must be registered to see the links

using the LCM sampler, 7 steps (4-8 works), 1.75 cfg (1-2 works), 0.5 denoising strength (can vary this number to stick closer to the original, or give SD more freedom to deviate) and 768x1152 resolution, and also using ADetailer for face improvement.

The Lightning models work really fast, I can generate an image in a few seconds, which makes this much more usable when playing the game. Regular SDXL models that require 20-30 steps take significantly longer, slowing down the flow. Some other models I had good luck with were

You must be registered to see the links

,

You must be registered to see the links

and

You must be registered to see the links

My focus also was mostly on realistic images - have not played much with more stylized models, anime, etc.

If anyone is interested in the modified GenAI mod files, I can post them. It's a bit hacked together right now (was just messing with it for personal use), but the changes are relatively simple - anyone who got the a1111 API to work in the game, and has any basic python knowledge could easily get this going as well.

It could probably be cleaned up and incorporated into the mod at some point (maybe the author of the mod can consider adding img2img as an option)

what do the nude models look like?

 

[ycbelane]

That's a good idea. I like to keep my models sorted, so I put them in folders but GenAI either escapes or doesn't escape the slash, so stable diffusion doesn't recognize the path to the model. Makes switching models easier for me.

Just a tip, especially if you create a lot of images: Change settings in webui to save images as jpeg instead of png (WebUI: Settings -> Search -> File format for images), my GenAI folder is 400MB big and that's from banging like 5 chars.
I'm using different triggers, like more positions (anal piledriver, skull fuck etc.), cum stains, or changing the background to the current location. This causes an exponential increase in generated images. Not sure how many images your modification generate.

Yes, I keep my models in subfolders as well (there are so many.... :) - so that was the issue I ran into (with the handling of the slash) - probably not hard to fix, but didn't seem important enough to spend time on yet, as I was already mainly testing different models in webui (copied a prompt from the game manually, and then tried different models/settings and prompt variations in there, to speed things up - only once I had a good working setup I moved it back into the game for actual playing)

Good advice on the jpg generation - I already keep autosaving off in webui, but switching the format to jpg there also affects the API results, and the GenAI mod is saving the generations to the "generated_images" folder, so it does save space (the jpg's are like 1/10th the size)

It's definitely a tradeoff in how dynamic and detailed the prompt is - the more dynamic and detailed, the more generations, which takes more time and space... my version so far doesn't add much more to the original, so it's not too crazy, but I have been experimenting with the prompting trying to find what is worth including and what isn't. One fundamental issue is how different models react to prompts differently, and doing too much microtuning for a specific model then all blows up when I change to different model. It would probably require some more sophisticated framework to allow handling this properly, where the prompt building could be customized for each model - but obviously that's not trivial, so for now I was just keeping the prompt simple and only addressing most glaring things.

 

[Arguendo]

Tried downloading and installing the latest Beta version (2025.07-betaVTMod4.0.31) via the Git auto function. I keep getting an exception error at start up every time. Have tried re-running the bat and remove rpyc, but no change.
Looks like there's an attribute missing for highlighting text in the menu somewhere. Know of a workaround for this?

Happens right after talking about meeting Stephanie at the lab.

Spoiler

You don't have permission to view the spoiler content. Log in or register now.

 

[ibnarabi]

Tried downloading and installing the latest Beta version (2025.07-betaVTMod4.0.31) via the Git auto function. I keep getting an exception error at start up every time. Have tried re-running the bat and remove rpyc, but no change.
Looks like there's an attribute missing for highlighting text in the menu somewhere. Know of a workaround for this?

Happens right after talking about meeting Stephanie at the lab.

Spoiler

You don't have permission to view the spoiler content. Log in or register now.

You need to roll back before that 'highlight_green' commit, it breaks things.

You must be registered to see the links

 

[ycbelane]

what do the nude models look like?

The Lightning models are not as good at fully nude generations, and the poses in the original game are kind of limiting as well, so the nude and sex positions are a little bland (not that the original renders in game are much better though :) - clothed and partially clothed generations are better (the AI stuff is generally really good at titillation and suggestive stuff, actual explicit and realistic porn requires specialized models and more prompting, which is more work to get dialed in)

Here are some examples:





Sometimes the generation does get a little weird (in some poses the ADetailer face fixing fucks up, or the hands/fingers can be off, typical AI stuff..) but in general it's not really much worse than the original images in the game, and sometimes way better. It gets the job done, and the speed/quality tradeoff with the lightning models I think is generally worth it.

Examples of when it's not as good or straight up fails:

(look at the feet on the last one :)

Using regular (non-lightning) models trained on porn and text2img with better scene descriptions would likely yield better results, but the img2img stuff has a place I think, adds to the variety somewhere between the original renders, and full on AI generated images.

 

[rb813]

Here are some examples:

I'm surprised at how similar everything below the neck is to the original LR2 assets. That can't be just based on the general description, right? The AI must have been giving those inputs at some point to "learn" about them, I would expect?

 

[ycbelane]

I'm surprised at how similar everything below the neck is to the original LR2 assets. That can't be just based on the general description, right? The AI must have been giving those inputs at some point to "learn" about them, I would expect?

It's img2img - basically, instead of starting with 100% random noise like regular AI text2img generation, it takes the input image (original one from the game in this case), then some amount of random noise (how much is controlled by a parameter) is applied to the source image to create a more vague version of it, then the generation is allowed to resume from that point to remove that noise. That means that the input image can very strongly influence the final output. The AI model has been trained on a huge amount of real life photos of people, so it has a very good general model of what a variety of human bodies look like in various lighting conditions and poses. It can apply that knowledge to the vague input and refine it back to a realistic looking full image. Essentially the AI can imagine what a more real life like version of that game image would look like from what it knows, even though it was never trained on any images related to the game directly.

The prompt description in this case can help guide the generation. Normally, when the generation starts with 100% noise, the resulting image is all formed based on the prompt. However, when starting from a source image, the prompt can still push the model to interpret the noise a certain way or in specific context, while still staying close visually to the original.

Also, since the source images in the game are already semi-realistic, the full-realistic versions made by AI are pretty similar. The AI can go a lot further than that, using for example cartoon images or drawings as source, and still generate equivalent realistic versions, or vice versa, take a real photo, and generate a cartoon version.

Here is an example of just some random clipart image of a girl fed as input to the same img2img model as I used above, and on the right is what it generates:



So as you can see, the model needs very little "human like" shape to work with, and can turn it into a realistic image. Also, I had to add "girl holding a martini" in the prompt to get it to generate the martini glass, without that prompt, it just ignored it and made her hand empty, so that's an example of how the prompt is still useful.

Also here are 2 other variations based on the same source image above, one with lower denoising strength (forcing it to follow the source more closely, resulting in still mostly cartoony image, but with somewhat more realistic human proportions than the original) and one with higher (allowing it to be looser, in this case it decided to change the pose and even the dress a bit):

or I can specify her race in the prompt to further guide the generation:

It's honestly pretty crazy what generative AI can do these days, especially given that it can generate one of these images with a 6GB model in about 3 seconds on my local computer (with a 4090 I admit, but still - I was lucky enough to buy one when they first came out before the AI stuff really exploded, and they cost half as much as they do now)

 

[themagiman]

The Lightning models are not as good at fully nude generations, and the poses in the original game are kind of limiting as well, so the nude and sex positions are a little bland (not that the original renders in game are much better though :) - clothed and partially clothed generations are better (the AI stuff is generally really good at titillation and suggestive stuff, actual explicit and realistic porn requires specialized models and more prompting, which is more work to get dialed in)

Here are some examples:

View attachment 4816836 View attachment 4816837 View attachment 4816838 View attachment 4816891 View attachment 4816840 View attachment 4816841 View attachment 4816843

View attachment 4816901 View attachment 4816902 View attachment 4816903 View attachment 4816904

Sometimes the generation does get a little weird (in some poses the ADetailer face fixing fucks up, or the hands/fingers can be off, typical AI stuff..) but in general it's not really much worse than the original images in the game, and sometimes way better. It gets the job done, and the speed/quality tradeoff with the lightning models I think is generally worth it.

Examples of when it's not as good or straight up fails:

View attachment 4816860 View attachment 4816861 View attachment 4816876 View attachment 4816908 (look at the feet on the last one :)

Using regular (non-lightning) models trained on porn and text2img with better scene descriptions would likely yield better results, but the img2img stuff has a place I think, adds to the variety somewhere between the original renders, and full on AI generated images.

Still looks amazing vs current images. I wish it could get the chokers right with the words and not create errors like you showed.

 

[Jerokor]

Hello everyone, i happened to stumble across some inconsistencies during gameplay in version 2025.04: certain scenes do not match the girls preferences. Especially the scenes after sleepover need to adapt to the opinion system. I managed to acquire a girlfriend ho hates vaginal sex, but in the morning she used her vagina with passion.
I also cannot pick the orifice for the morning dicking and in some scenes, where she takes control, she also chooses vaginal. That's quite odd.

Is this going to change so the opinions/preferences are way more considered during initiative scenes or in the morning?

 

[rb813]

It's honestly pretty crazy what generative AI can do these days, especially given that it can generate one of these images with a 6GB model in about 3 seconds on my local computer (with a 4090 I admit, but still - I was lucky enough to buy one when they first came out before the AI stuff really exploded, and they cost half as much as they do now)

That is really interesting, thank you for sharing!

 

[a1fox3]

Hello everyone, i happened to stumble across some inconsistencies during gameplay in version 2025.04: certain scenes do not match the girls preferences. Especially the scenes after sleepover need to adapt to the opinion system. I managed to acquire a girlfriend ho hates vaginal sex, but in the morning she used her vagina with passion.
I also cannot pick the orifice for the morning dicking and in some scenes, where she takes control, she also chooses vaginal. That's quite odd.

Is this going to change so the opinions/preferences are way more considered during initiative scenes or in the morning?

Some girls like Sarah and a few others have story sex that will not be changed to coincide with opinions.

 

[Darkmind]

As a Mac user who is often just using Wine/Crossover: Does the AI Gen mod connect to Stable Diffusion using a IPC call or a local URL?

(Basically: Do I have any hope of being able to run this without completely rebuilding everything (URL) or is this all dependent on everything being in the same Windows environment (IPC)?)

 

[osmcuser132]

As a Mac user who is often just using Wine/Crossover: Does the AI Gen mod connect to Stable Diffusion using a IPC call or a local URL?

(Basically: Do I have any hope of being able to run this without completely rebuilding everything (URL) or is this all dependent on everything being in the same Windows environment (IPC)?)

I can run it on linux, it opens and listens on a TCP socket (port 7860)

 

[ycbelane]

As a Mac user who is often just using Wine/Crossover: Does the AI Gen mod connect to Stable Diffusion using a IPC call or a local URL?

(Basically: Do I have any hope of being able to run this without completely rebuilding everything (URL) or is this all dependent on everything being in the same Windows environment (IPC)?)

The mod uses HTTP to connect and communicate with your stable diffusion/a1111 instance using the python requests module - currently it's hardcoded to connect to

You must be registered to see the links

but you could easily modify the code to change it if you have SD running on a different host or a VM with its own IP or whatever custom situation you might have.

 

[ycbelane]

Still looks amazing vs current images. I wish it could get the chokers right with the words and not create errors like you showed.

Yeah, getting small details like this right is not straight forward, at least currently, although there are models now that can do text pretty well (not sure if that tiny) However there are ways to make it work - it's possible to train special small models called LoRA's which can be used to microtune a full model to introduce a new named concept. This can be done with relatively few source images - for example generating a variety of LR2R renders with the collars with different words, and feeding them to LoRA training with appropriate tags. The generations then could be guided to include these collars using those tags in the prompt.

In theory this shouldn't be too difficult, although I have not trained a LoRA myself yet. It's definitely a bit more work than the minimal changes I had to make to the mod so far to make img2img work in the first place.

 

[Darkmind]

I can run it on linux, it opens and listens on a TCP socket (port 7860)

The mod uses HTTP to connect and communicate with your stable diffusion/a1111 instance using the python requests module - currently it's hardcoded to connect to

You must be registered to see the links

but you could easily modify the code to change it if you have SD running on a different host or a VM with its own IP or whatever custom situation you might have.

Perfect, thanks. Sounds like it should be fairly straightforward then, I'll have to try it out.