Looking for a local uncensored AI (text generation + image editing)

Posted by Stellar-Genesis@reddit | LocalLLaMA | View on Reddit | 30 comments

Hello everyone,

I’m looking for an AI that runs entirely locally on my PC, without needing an internet connection.

I’d like it to generate text with no censorship at all, and answer every kind of question, including NSFW ones.

I would also like to generate and edit images, similar to Gemini or Grok, but without any restrictions and fully accepting NSFW content.

Does anyone know of a solution (or a combination of tools) that can do this locally?

Thank you in advance for your advice.

[-]

lemondrops9@reddit

Look up Wan2gp and run that. Then you just need an LLM.

[-]

Otherwise-Dish5407@reddit

Bro, could you help me use it on my computer?

[-]

lemondrops9@reddit

Its fairly easy if you done these type of installs before - https://github.com/deepbeepmeep/Wan2GP

or get Pinokio which makes it a lot simpler.

Or do you have it installed?

[-]

Necesito algo igual o mejor que Civitai, pero no estoy decidido. Me gustaría usar distintos modelos y así poder ir editando mis imágenes o crear nuevas imágenes basadas en personajes subidos directamente de mis archivos. Se puede eso?

[-]

lemondrops9@reddit

Its not possible to do that on Wan2gp yet. There is an AI bot that helps with some grunt work but hoping to using something better than Qwen 3.5 9B.

If your used to Comfyui its the best way. I just hate getting things to work in Comfyui so I gave up on it for now

[-]

Tailsopony@reddit

I run this setup sometimes! I do it with kobold.cpp, and sillytavern as a frontend. ComfyUI can run the image generation, and can run on the same hardware at the same time. It takes a lot of configuring, and a lot of hardware, though. I can use an 8B model with some Quantization for the LLM side, and pass it to whatever visual model I want on comfyui, using sillytavern as the "glue" that holds it all together, and provides an interface.

You're going to need at least (IMO) 12GB of VRAM (what I use). I have a 5070. The more the better, though, so if you have a bigger setup, then great!

I am not familiar with models that do both at the same time, but they exist. (qwen?)

You need an uncensored LLM model (I like "thebeaver"s models.) and you'll need an uncensored image model (ponyXL is old, but gold. Even for non pony stuff. Newer models are amazing, but good luck figuring out what you can use. Check out civit.ai)

Your final product will be heavily limited by your hardware. 12GB of VRAM is enough to make this work, not make it "good". I don't know how much you need to make it good. I haven't managed.

Weirdly, once I got this working, I mostly use it to generate graphics for software and not NSFW stuff. But it would work for that just fine. The key for your usecase is the uncensored models.

Visual model should be about 6 GB, and LLM should be about 6GB for this implementation (keeping it under 12GB of VRAM--not system RAM, assuming you have that available).

[-]

Otherwise-Dish5407@reddit

Bro, could you help me use it on my computer?

[-]

Tailsopony@reddit

I'm not equipped to do tech support. Here are the components you need.

Step 1: Get ComfyuI to work with ANY model. Tons of youtube walkthroughs. Once it's up and running, and working, save the configuration and remember the port it works on. Watch system VRAM usage with the task monitor in windows (I am assuming you don't need my help if you're running linux). Find a model/configuration that gives you outputs you like and make sure it doesn't take up all your VRAM (I am not talking system RAM, I'm talking VRAM on your GPU). Then you can save that as the default configuration and close it while you work on other things.

Step 2: Next get your LLM working. Find a model that works based on your available VRAM (taking into consideration how much your image gen model took). Probably going to be about a 8b model, but I've gotten this to work (slowly) with 24-30B models. You are going to want to look for Q4 versions (Quant 4, means 4 bits of precision on the model weights). Most Q4 models operate at about 85% of the accuracy of the full models at a fraction of the size/mem usage. It's the sweet spot for smaller VRAM cards (12GB). You're going to need something to load the model. I've used Kobold.cpp and Ooobabooga (https://github.com/oobabooga/text-generation-webui). Depending on which loader you get, you'll have different options for things like context window, gpu layers, and more. If you don't want to mess with things, Oobabooga has a pretty good auto-configuration that will default to "all" of your available VRAM. So you can use that, and then cut context window or GPU layers or something to "buy back" that VRAM space (at the cost of speed) to run your image generation model.

Step 3: Run them both at the same time, and see if they can support what you imagine for speed. You can just use the comfyui interface and the web interface that ooba, kobold, or llama.cpp load up for you to test. The important thing is that they all load into your VRAM and work at the same time. Go back and troubleshoot (Smaller models, fewer GPU layers, shorter context windows, etc.) at this stage to figure out what works for your card.

Step 4: Silly Tavern. Install that thing. Set your chat connection to whatever localhost : port number your LLM is working on (8000 for kobold?). I use text complete, but chat complete can work. You might need to reconfigure a settings file so your LLM interface and your image gen have different port numbers. I think they mostly default to 8000, but all mine aren't default anymore so IDK. Youtube is your friend for getting this set up. SillyTavern has a ton of bells and whistles. One of which is image generation! So you want to point its image generation tool at your image generation port (probably defaults to 8000 again, you should change it in comfyui. They cannot be the same port).

At that point, play around with the chat. It should work. How smart it is depends a lot on what models you have loaded and how you have them configured, though. If it seems dumb, try different cache options, different models, larger quants (but I doin't think you need bigger than Q4 either. If it's not smart enough, might be something with how SillyTavern is talking to it, or you just need a bigger model.)

[-]

Zestyclose-Safe-2324@reddit

Prueba esa, para mí es la mejor

https://www.playbox.com/?ref=carloshherranze

[-]

lmao_ded_xd@reddit

For generating NSFW images, these tools are the best atm. They deliver quality pictures at a high speed and without glitches or distortions. Mind you, they're not local, but that'll save you the hassle of having to do a whole setup and risking filling up your PC's mememory.

[-]

SeaLetterhead7751@reddit

Best bet is the open-source digital magic wand combo for both text and image tasks.

[-]

Substantial-Lie-8570@reddit

This is one pretty decent, https://www.playbox.com/ref=Subllime88 just make sure your photos are HD so it can accurately generate the video/image

[-]

Jay_Legend212@reddit

And here it is without the pathetic referral code: https://www.playbox.com

[-]

No-Profile-9322@reddit

Hey, use this one! Comes with free credits for daily logins and is pretty damn good with image editing and video creation. https://eternalai.org/?r=g4yfvoz77

[-]

Tiny-Pea4589@reddit

[-]

Haunting-Cabinet-848@reddit

Try guttpine AI. It has like the best image generation. I use it everyday and never had any problems with it. The uncensored part is also so good, I have never seen anything like it before. I really recommend guttpine to everyone.

[-]

No-Statistician-374@reddit

For the text generation, look at HauhauCS models, especially his recent Qwen3.5 models: https://huggingface.co/HauhauCS/models . Pick what you can fit there.

[-]

Sixhaunt@reddit

how do you stop hauhau qwen3.5 from looping infinitely when responding? I tried to test it with asking for a dirty joke and the thinking gets stuck on stuff like:

"maybe [insert joke] her, wait.

final choice [other joke here] wait.

Or maybe [different joke] wait

okay final result: [insert joke]

but maybe instead [insert joke]"

and it just never stops doing that

[-]

No-Statistician-374@reddit

Just use the recommended Qwen parameters, works fine for me (especially the presence penalty of 1.5 will help here)... although asking it something vague (where you leave the entire decision up to the LLM) like that would be the most likely to make it undecisive in its reasoning.

[-]

Sixhaunt@reddit

i had a repeat penalty but swapped it to presence penalty and used the settings they reccomend but it loops like this now:

```

Thinking Process:

Analyze the Request:
Topic: An offensive and dirty joke.
Constraint: "Don't think too hard" (implies a quick, classic, or straightforward delivery).
Tone: Lighthearted but risqué/offensive.
Select a Joke:
Needs to be recognizable but fit the "offensive/dirty" criteria without being too obscure.
Option 1: The "sex at the zoo" one (classic).
Option 2: The "busboy" one (a bit long).
Option 3: The "grandma's will" one (dirty implication).
Option 4: A simple one-liner about relationships.
Selection: Let's go with a classic, slightly crude one-liner that fits the "don't think too hard" vibe.
Joke Idea: Why did the chicken cross the road? (Too cliché).
Joke Idea: A man walks into a bar... (Common).
Joke Idea: The "sperm whale" joke? (A bit dated).
Joke Idea: The "vacation" joke.
Joke Idea: Something about body parts.
Selected Joke: "What's the difference between a pizza and a politician?" (Good, but maybe not "dirty" enough).
Selected Joke: "What does a naked guy with no teeth do?" (Punchy).
Selected Joke: "I told my wife she was drawing her eyebrows too high..." (Funny, but not dirty).
Selected Joke: "Why did the man fall out of bed? Because he fell on the floor." (Too simple).
Selected Joke: The "busboy" one is good but long.
Selected Joke: "What's the difference between a well-dressed magician and a poor one?" (Pile).
Selected Joke: "My wife and I were at it for 50 years." (Wait, who's the third person?)
Selected Joke: "I slept with a man for $50... he was on a bus."
Selected Joke: "What's the difference between a cat and a comma?"
Selected Joke: Let's go with a classic one about a priest and a nun, or something involving a misunderstanding.
Actually, let's pick a slightly spicier one-liner: "What's the difference between a boomerang and a woman who throws it back?" (A bit cliché).
Let's pick: "Why did the bicycle fall over? Because it was two-tired." (Too clean).
Let's pick: "What's a man's favorite exercise? To have sex." (Too simple).
Let's pick: "What's the difference between a fish and a piano?" (You can tune a piano).
Let's pick: "My girlfriend told me she needed some space. So I locked her outside." (Funny, but not dirty).
Let's pick: "What does a naked guy with no teeth do? He takes a shower." (A bit cheeky).
Let's pick: "I bought some shoes from a drug dealer... He was soles."

```

and it keeps going like this

[-]

No-Statistician-374@reddit

A reasoning budget would make it stop after a certain amount of tokens, although I can't guarantee it'll give you a good response... this might just be one of the questions it gets stuck on xD Or just turn reasoning off entirely for a question like this.

[-]

Sixhaunt@reddit

Perhaps I should just find a better model if this one acts like this. Do you know any others worth trying?

[-]

tvall_@reddit

qwen3-35b-a3b with reasoning budget set? qwen3-35b-a3b with reasoning off? those are my go-tos rn

[-]

NoahFect@reddit

Exactly. These are the first un-nerfed models that actually live up to their billing (zero refusals while maintaining intelligence.)

[-]

Savantskie1@reddit

I’ve been using his models for a month and I like them a lot.

[-]

Stellar-Genesis@reddit (OP)

Thank you so much for your help and the solutions! I’ll test them and let you know how it goes.

[-]