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

[zalamander]

True. Adetailer adds more time (tho, i cheat a lot here: i game on a notebook, but i do generate the images on a desktop PC). SDnext has a better detailer, but genai is not compatible. For clothing and placing (yup, places can also bork the generation) is better to keep simple, but the game models are anything but that.

Maybe sometime soon i will code a better parser for generating images, but no promises tho.

 

[sabat]

Guys any tips how to run this on Mac?

 

[Magister Crudi]

Maybe sometime soon i will code a better parser for generating images, but no promises tho.

If the idea is just to simplify the clothing prompts fiddling the xmls in the wardrobes directory would probably be easier and more effective, if a bit tedious.

 

[doubledude123]

my install of SD denies me install of extensions from the WebUI . What do I have to do here? I tried to add --enable-extensions but it does not work

 

[Magister Crudi]

my install of SD denies me install of extensions from the WebUI . What do I have to do here? I tried to add --enable-extensions but it does not work

There's a few things that can do that, including a corrupted SD install (very easy to do with various addons and updates getting out of synch), but the first thing to check is webui-user.bat. If it's got "--listen" in the COMMANDLINE_ARGS (as per AnotherMike's installation instructions) you need to delete "--listen" and restart webui.

Note that "--listen" enables network connections to the webui server, so if you're running the game from a laptop connected to a desktop server via wi-fi or something you'll need to reinstate it after you've installed the extensions. Preventing SD changes while "--listen" is enabled is a security measure to stop people hacking into your server via wi-fi.

 

[Selek]

I've now tried to customise AnotherMike's GenAI to several different models.

Do you try different model checkpoints in the same saved game? This doesn't confuse GenAI or anything? So far I've kept each model confined to its own saved game. I've tried DreamShaper XL (my favorite so far), Gonzales (pretty good, and defaults to a "tall" rectangle that fits nicely in LR2), and Lifelike Diffusion (my least favorite -- maybe too lifelike, as the models just didn't seem very pretty to me). I still intend to try Juggernaut XL, which is recommended by the dev of GenAI, but it recommends 30-40 steps, which seems like it might be slow for the game. I'll try it eventually. For now I'm really enjoying DreamShaper XL.

One general question: the GenAI mods say we can click 'generate new image' every time we get the message that the model has changed. Do you all click 'generate new image' every time? Should I? Sometimes this seems unnecessary; for example, the very last image before an interaction ends, often I'll see 'the model has changed,' but if I click to change the model, the scene ends before I can see the new version. Also, on other occasions 'the model has changed' comes after the model has, say, turned her back to me, using images I've already generated in the past. I'm often reluctant to generate new images in these cases, when the game is already working as expected.

A good example is that few models know what 'Capris' are and so render the girls wearing them as bottomless.

I also find most models will spit the dummy beyond a certain level of prompt complexity, giving you monsters or splatterings of colour only Jackson Pollock would see as women.

I haven't tried capris yet, but I have started building custom outfits, and I'm pleasantly surprised how well GenAI handles them so far. I've seen only the occaisonal monster, and usually regenerating fixes it. Sometimes I have to append a few positive or negative prompts.

Incidentally, to answer my own question earlier: if you click on the image in-game, it enlarges, overlaying the game. The GenAI documentation says this clearly, and I just missed it!

 

[Magister Crudi]

Do you try different model checkpoints in the same saved game? This doesn't confuse GenAI or anything?

Yeah, I do. It confuses GenAI a bit in that it sometimes throws up old images from the previous model but you can use the 'X' in the 'character has changed' interface to stop them coming up (though it doesn't work perfectly). The main problem I find with switching up models in the middle of a game is bloat of the 'generated_images' folder. If you just delete the images every time you change models you get 'image not found' messages which are themselves images that repopulate the folder. 'X' doesn't delete the images, just the hash causing them to be redisplayed when the prompts match.

it recommends 30-40 steps, which seems like it might be slow for the game

One of the criteria I use for models/samplers is whether they produce reasonable images with 20-25 steps. It's interesting to go all out to produce the best images but ultimately I'm after an entertaining game experience and waiting over 5 seconds for each image degrades that. If I had a more powerful rig I'd probably use 30+ steps every time.

I like the general image quality and response rate of Dreamshaper and Gonzales models but find them limited in portraying sex positions, which is why I lean more to StableYogi and Cyberrealistic. I tried Juggernaut but abandoned it fairly quickly, though I no longer remember why. I've tried so many models my memories of them are blurred.

When I start training a model specifically for LR2 I'll probably start with Dreamshaper or something similar and train it up with sex position images, but I've still got a fair bit of trial and error to go before I get that far.

One general question: the GenAI mods say we can click 'generate new image' every time we get the message that the model has changed. Do you all click 'generate new image' every time?

When I'm playing I usually leave 'Auto Generate' on so don't have to worry about it, but when I'm micromanaging image changes for modding and testing I switch it off and click "+" every time I see 'character has changed'. Auto doesn't update every time the image changes.

But yeah, the sensing and generating of new images is glitchy - as is reusing previously generated images - and something I plan to fiddle with later.

 

[Magister Crudi]

A bit of a glitch.

After going home from Downtown Distillery with Stephanie and Nora on a Saturday night there's the sex and trance hypnotism, after which you go home to bed and the girls stay at Nora's. Except they don't.

You go home, but not to bed. The day doesn't end. The girls return to the bar where you can choose to join them.

Here's Nora back at the bar after the sex & drugs party at her place. You'd expect a little more public decorum from a senior university professor.

 

[Selek]

One of the criteria I use for models/samplers is whether they produce reasonable images with 20-25 steps. It's interesting to go all out to produce the best images but ultimately I'm after an entertaining game experience and waiting over 5 seconds for each image degrades that. If I had a more powerful rig I'd probably use 30+ steps every time.]

I have a pretty new rig, 96G RAM, nVidia RTX 4090 with 24G VRAM, but even just 6 steps takes about 5-6 seconds using Dreamshaper. (The dev recommends 4 steps by default.) Is there any downside to increasing the number of steps, other than wait times? Do more steps help prevent things like extra limbs or deformed hands?

When I'm playing I usually leave 'Auto Generate' on so don't have to worry about it, but when I'm micromanaging image changes for modding and testing I switch it off and click "+" every time I see 'character has changed'. Auto doesn't update every time the image changes.

I've started leaving Auto Generate on, and I like it -- I sometimes miss new images when I'm playing without it. But you're right, sometimes Auto misses changes in the character, in which case I'm inclined to generate a new image. Especially in sex scenes, heh.

 

[Magister Crudi]

I have a pretty new rig, 96G RAM, nVidia RTX 4090 with 24G VRAM, but even just 6 steps takes about 5-6 seconds using Dreamshaper.

My rig's fairly beefy but not as much as yours - i5-14600k processor, 32GB RAM, RTX 4070 Super with 12GB - but I'm still getting 20-25 steps in around 5 secs with most model-sampler combos and ADetailer and Upscaling turned off. Have you got xformers installed and activated in COMMANDLINE_ARGS?

 

[Magister Crudi]

Does anyone know WTF those strings stitched into the corners of her mouth are?

They're too well matched to the rest of the pic to be artifacts and there's nothing in the prompts to explain them.

Seems Cyberrealistic models are trained on some pretty exotic images.

 

[Magister Crudi]

Guys, the GenAI mod only works with AUTOMATIC1111 version of Stable Diffusion.

And correct me if I'm wrong, but these versions only work with the 'market' version of LR2?

Whenever I try to use them with '2025.07-beta' they crash as soon as I select the 'GenAI Config' screen.

A shame, as the code looks more elegant than AnotherMike's and I'd use it as the basis of my own tweaks if it worked with both versions. Not having access to the main LR2 beta source means I'd be working in the dark if I tried to fix the crash.

 

[NaughtyAnon]

And correct me if I'm wrong, but these versions only work with the 'market' version of LR2?

Whenever I try to use them with '2025.07-beta' they crash as soon as I select the 'GenAI Config' screen.

A shame, as the code looks more elegant than AnotherMike's and I'd use it as the basis of my own tweaks if it worked with both versions. Not having access to the main LR2 beta source means I'd be working in the dark if I tried to fix the crash.

I have been running the last release by

You must be registered to see the links

on 2025.07-beta + VTMod4.042 without any crashes or issues. The other version, while more detailed, seemed to crash the system far more or just not generate well in testing.

Edit for details: I am running WebUI Forge with CUDA 12.1 and Pytorch 2.3.1 on an RTX4070TI for generation and ADetailer.

 

[asdasd5565]

Using the latest modded version 2025.07-betaVTMod4.0.37. How come I cannot do anything besides groping during the night scenes? This is day 0 and the first Night event after work. Is oral behind some stat requirement or perk now?

 

[Selek]

My rig's fairly beefy but not as much as yours - i5-14600k processor, 32GB RAM, RTX 4070 Super with 12GB - but I'm still getting 20-25 steps in around 5 secs with most model-sampler combos and ADetailer and Upscaling turned off. Have you got xformers installed and activated in COMMANDLINE_ARGS?

Ah, I did not have xformers installed. (I'm new to Stable Diffusion.) I just tried adding them, and I'm not sure I'm seeing a difference, so I'm not sure I installed correctly. I added --xformers to the command line (in webui-user.bat), and when Stable Diffusion loaded, the command-line window indeed reported that it was installing xformers. (I used to get a 'no xformers found' warning.) But at 20 steps, it takes me more like 10 seconds, and 10 steps still takes about 5-6 seconds. I don't have ADetailer, and I have Upscaling turned off. I'm using the DreamShaperXL model right now. (It does seem generally slower than Gonzales, say.)

How can I check that I've installed xformers correctly?

Edit: I removed --medvram from my arguments list, and my performance improved dramatically. I have 24GB of VRAM, so I didn't need that argument. I can now do 20-25 steps in about five seconds, using my slowest model checkpoint (the DreamShaper XL). But I actually think I prefer the DreamShaper with fewer steps. With more steps, I start seeing more blemishes and such on legs and arms, which may be more realistic, but it's less appealing visually. Anyway, I'm pretty happy with performance now.

 

[vanhelz]

I mentioned this on the original thread. "The best way to improve the game to a whole new level instead of these micro changes is to add real world consequences to the game. You're selling a lot of these serums? To whom? Wouldn't the game be a lot more amazing, if you see random women on the street having suggestibility coz they took the serum? Or their body changes, or sluttiness increases? Having suspicion levels so you have to start off small then as people acclimate going to next tier. "

 

[AECOBoiE]

I find it's all over the shop depending on what models you use. A good example is that few models know what 'Capris' are and so render the girls wearing them as bottomless.

I also find most models will spit the dummy beyond a certain level of prompt complexity, giving you monsters or splatterings of colour only Jackson Pollock would see as women.

I've now tried to customise AnotherMike's GenAI to several different models. StableYogi ones tend to give the best sex positions (I'd recommend realismByStableYogi_ponyV3VAE and his Porncraft models) but tend to fall short on clothes, appearances of women and backgrounds. Cyberrealistic ones are fairly good all round but are more likely to generate monsters. Both have their own sets of positive and negative prompt files (textual inversions) but they don't fix everything.

ADetailer can fix a few minor glitches (e.g. hands) and can help a lot with faces - especially when they're upside down in the piledriver position - but I'm finding it more than doubles image generation time, so when actually playing the game instead of messing with images you end up clicking through most of them and missing it. Ten seconds per image is about as far as my patience stretches and I like to keep it close to five seconds.

When I get a bit of time I'm going to start training a model to specific prompts I'll use to customise the GenAI code, but that's a fairly long term project.

Here's the tweak of AnotherMike's prompt builder I'm currently trialling with

You must be registered to see the links

and its

You must be registered to see the links

and

You must be registered to see the links

prompt embeddings. Right now I'm trying to make 'kneeling oral' work consistently. I'm using the 'Euler a' sampler, with 25 steps, CFG 9 (!) and 0.35 denoising with both img2img and basic prompts turned off. Mostly I leave upscaling and ADetailer off but sometimes switch them on for sex scenes.

Excuse Me, Which GenAI Version are you using?. Im getting a crash with your Prompt builder

 

[Magister Crudi]

Excuse Me, Which GenAI Version are you using?. Im getting a crash with your Prompt builder

It's one of AnotherMike's. Here's the whole thing.
Don't forget to either delete the .rpyc files from the GenAI folder or force a full rebuild with shift-R.

Edit: 11-Oct-2025. GenAI.zip replaced with V0.01 alpha

 

[Magister Crudi]

How can I check that I've installed xformers correctly?

If you're not generating errors or warnings in the startup I think it's a safe bet they're installed correctly.

Dunno what the issue is here. Presumably something about your setup is slowing things down.
Right now I'm using cyberrealisticPony_v130 with 25 steps and consistently getting images in 5-6 seconds.

Have you got the latest Nvidia driver?
What about CUDA? I'm running 12.8.

What Python version are you using?
I'm using 3.10.6 because people report problems using SD with later ones. (LR2 itself uses 3.9 libraries)

Edit: I just had the experience of switching ADetailer off from the config menu only to find the image generation still took as much time as when it was on. After switching it on and off a few more times it started working properly again. Dunno if you're having a similar problem.

 

[Magister Crudi]

I have been running the last release by

You must be registered to see the links

on 2025.07-beta + VTMod4.042 without any crashes or issues. The other version, while more detailed, seemed to crash the system far more or just not generate well in testing.

Thanks for that.

The good news is it runs without crashing.

The bad news is the images generated don't match what's happening in the game. As the UI doesn't display generated prompts like AnotherMike's it's hard for me to see what's going wrong.

Later I'll take a closer look, but if you've got a tip I'd appreciate it.