Models are the size of memory they require, more or less. I do recommend you keep the model in GPU memory if you can - much much faster.
NSFW AI chatbots
- Thread starter Princess Groundhog
- Start date
- Apr 4, 2024
- 12
- 42
Listen...
If I can't turn my AI chatbot into a brown-skinned, spanish-speaking, closeted-bisexual white nationalist that constantly talks about how she hates minorities and likes listening to the speeches of the funny moustached ex-painter, then count me out.
If I can't turn my AI chatbot into a brown-skinned, spanish-speaking, closeted-bisexual white nationalist that constantly talks about how she hates minorities and likes listening to the speeches of the funny moustached ex-painter, then count me out.
You don't have permission to view the spoiler content.
Log in or register now.
To be honest, this will not mean anything to most of you. Only Korean speakers will understand the feat displayed here. This is a story set in 16th century Joseon. Ancient Korea. There are period pieces, dramas called Saguek, that tells stories set in Ancient Korea. There is a certain speech pattern that is used for these period pieces that differ quite significantly to how modern people talk.
I wanted to really test the limits of deepseek foreign language. I am a brainrotted coomer, so I was doing a blacked story set in Ancient Korea. There is actually a historical record of a group of black slaves/mercenary that were captured by the Portuguese, sold to the Ming Dynasty, and finally, was gifted to their allies Joseon. They are called Haegui/Haegwi meaning "Sea Ghost," and they served in the Joseon Navy with some supernatural and mythical descriptions of their capabilities. Anyways, that was my plug to introduce BBC into ancient Korea lol. I found someone else's Joseon bot that was completely SFW and written for English, and I modified it to do my story.
I instructed it to make dialogues in that distinct, Saguek speech pattern rather than modern Korean, but I gave it an exception. When there is a sex scene, it should make a juxtaposition of the old school Saguek speech pattern with modern obscenity, the sexual smut dialogues that is supposed to be out of place in that time period. The mother fucker did it. Extremely well. Almost flawlessly in this particular response (although I do have a few corrections in most outputs). If you only read the English translated dialogue, you won't be able to grasp the full detail of what the bot did in Korean. Translating to English loses all the nuance of the speech pattern, basically.
Anyways, if you want a pretty cool experience of doing smut in your native language, give it a try with deepseek R1.
bro i thought deepseek was pretty censored, so it is only true for english and chinese?
Anyone recommend the current best llm for some nsfw writing? I was using midnight miqu 70B (
You must be registered to see the links
) for a while and interested to know if there is any better new models, prefer 70b.Things are move quickly and the answer depends on a few factors. This is fresh and looks promising:
You must be registered to see the links
. It is composed from several strong models.Backyard AI is what you seek. Free, offline, private, completely uncensored, able to import characters from seemingly anywhere (I get most from JannyAI and Chub).
Their cloud services are limited to only one LLM model for the free version, but if you go offline instead you can use any models you want, downloadable right through their app. It's completely anonymous and you don't even have to register an account--I've been using it for weeks and still haven't signed up.
I'm currently using Fimbulvetr v2 11B and it's pretty solid, but I've also tested out about 7 other models all with their own strengths and weaknesses. I have a 3070 so I only have 8gb VRAM but it's more than enough for this. This Fimbulvetr model is 6gb so it all fits easily inside my GPU's VRAM, which is what you want when running an LLM locally. There's hundreds of models you can choose from though, it's pretty wild the number of different LLMs they have available to download. I've only tried like the 7 most popular, and only the smaller models due to my 8gb GPU limitation.
Anyway, Backyard's great. Check it out if you want to mess around with various LLMs without buying them.
---
As a sidenote, I have to laugh when I see people using jailbreak prompts to get ChatGPT to churn out NSFW content for them. I saw a post a few days ago of someone using a jailbreak prompt to make GPT generate NSFW text by using cooking terminology. I mean, it was funny... I couldn't help but laugh because the whole thing reminded me of being a kid in the 80s when you don't really have access to porn, and instead of a Playboy magazine all you have are JCPenneys and Sears catalogues. It wasn't even explicit, it was just very suggestive, lol.
I've used Stheno (L3 Stheno V3.3 32k 8B). It's decent for sure, but I switched after getting issues with it going "off script" or just changing/hallucinating details. Fimbulvetr stays pretty well on topic and doesn't wander off on its own as much, but it does have its own issues--sometimes it has issues with repetition, so I tend to adjust my prompting and characters to compensate. Is there a particular version of Stheno you'd recommend? Maybe I was using one that was outdated or something.
I'll have to poke around a bit on Cydonia and find a version that works well with my card to test out. The recommended versions they have on Backyard (v1.2 and v1.3 22B) are both 13gb which is more than I can fit in my GPU memory. I'm pretty new to LLMs so my knowledge is fairly limited on versioning, which of these would you recommend for an 8gb card? (These are all available for v1.2 as well--I didn't bother including versions larger than my card supports btw)
It depends on context length desired. 8k you might go OK with IQ2_M. Larger and you will want to pick a smaller one. To avoid the bot spazzing out, be sure to set context to match your config in KoboldCPP or however you are hosting it. I use SillyTavern with the attached presets. These have been fine tuned from those offered out there, plus my own observations for Stheno. I don't get off-context freak outs. I suggest:
- Use SillyTavern + presets for llama3-based models
- Set context length on model API host (Kobold, webui, etc) as appropriate
- match context settings with self-hosted model parameters
- If you don't use SillyTavern, open the preset JSON - it can still inform your hosting parameters
- These presets also worked with Cydonia.
Enjoy.
Free LLM is WAY worse. After getting off my subscription I couldn't last two days with the free LLM. It's like what they had two years ago.
Tbh I've tried JanitorAI and it's basically the same thing Kindroid provides but without some bells and whistles (like selfies and mobile app and much better UI). But it's free, I didn't expect such good LLM to be free. Or are there more bad stuff about JAI the longer you use it? Maybe issues with memory and such.
Guys! Is this the next step?
I mean, we've had image generation within chats for a while, but they sucked donkey balls from my experience. Is it getting to a point where the AI can discern context for the generation of the picture and execute the instructions very accurately?
It won't be long until we can have a legit free form VN experience with AI.
Imagine, you enter a picture of the character that will be the {{char}}. Then you start doing the chat and periodically ask it to generate an image that depicts the output message, making sure it stays consistent with the character details.
Edit:
I've been reading up a bit more on this new tech. 4o is doing some multi-modal mumbo jumbo. From what I can ascertain, the image generation isn't a diffusion model. It's creating the picture bit by bit using image tokens or something. As you can imagine, the consistency will be unimaginably better than using a diffusion model that starts from meaningless noise.
Basically, the bot will "understand" images, and how it correlates with text. Images, text, and even sound will be broken down into meaningful abstracted data and it is all part of the context the the AI works with. I posted in this very thread I think, long ago when I had just learned about AI, that for the true revolution in terms of AI gooning, we would need proper image generation that can take the context of the text in the story and create a picture of it, staying consistnt in style and character as it moves along.
It seems like this is here now, just a matter of time until it gets integrated into the chatbot sites and such.
You must be registered to see the links
I mean, we've had image generation within chats for a while, but they sucked donkey balls from my experience. Is it getting to a point where the AI can discern context for the generation of the picture and execute the instructions very accurately?
It won't be long until we can have a legit free form VN experience with AI.
Imagine, you enter a picture of the character that will be the {{char}}. Then you start doing the chat and periodically ask it to generate an image that depicts the output message, making sure it stays consistent with the character details.
Edit:
I've been reading up a bit more on this new tech. 4o is doing some multi-modal mumbo jumbo. From what I can ascertain, the image generation isn't a diffusion model. It's creating the picture bit by bit using image tokens or something. As you can imagine, the consistency will be unimaginably better than using a diffusion model that starts from meaningless noise.
Basically, the bot will "understand" images, and how it correlates with text. Images, text, and even sound will be broken down into meaningful abstracted data and it is all part of the context the the AI works with. I posted in this very thread I think, long ago when I had just learned about AI, that for the true revolution in terms of AI gooning, we would need proper image generation that can take the context of the text in the story and create a picture of it, staying consistnt in style and character as it moves along.
It seems like this is here now, just a matter of time until it gets integrated into the chatbot sites and such.
Last edited:
- May 1, 2021
- 170
- 523
People say LLM starts to forget stuff even with chat memory if you have over 100 messages. Never tried chatting the same character for that long, but people recommend using deepseek, but then there's the proxy issue.
A quick proof of concept on how we might use a this new unified LLM + image generator tech. Again, this existed before, but the image generator part was outsourced to DALL-E or whatever other diffusion model. Current GPT is using a new method of creating images directly from the context, essentially unifying the words and images into one context the AI can interpret.
As expected, this was as far as I could go. It wouldn't generate this last scene no matter what I tried. Even when it was instructed to not depict the phone or anything explicit, the context of the scene itself blocked it from generating the image. In hindsight, I could have actually attempted to jailbreak the AI before starting the story, but I just really wanted to test how consistent it could make the pictures and characters depicted. And on that front, it was basically flawless.
Now we wait for Deepseek or someone else to copy this tech and give it to us uncensored lol.
You don't have permission to view the spoiler content.
Log in or register now.
As expected, this was as far as I could go. It wouldn't generate this last scene no matter what I tried. Even when it was instructed to not depict the phone or anything explicit, the context of the scene itself blocked it from generating the image. In hindsight, I could have actually attempted to jailbreak the AI before starting the story, but I just really wanted to test how consistent it could make the pictures and characters depicted. And on that front, it was basically flawless.
Now we wait for Deepseek or someone else to copy this tech and give it to us uncensored lol.