← sources

Does a purely uncensored model exist - without a LOra?

source: reddit (reddit:StableDiffusion)published 2026-10-19 19:57ingested 2026-10-20 02:38

retained text

I am setting up my own inference server and going to lean on some Intel cards to do odd-jobs like STT, TTS and also image generation and editing. So far, I had noted Qwen Image 2.1 because I really like it's dual use-case and number of reference images that can be used - this aligns quite well with how I usually use image gen (and how I used to commission artists; long text, many refs).

But as I started to peek around the supported model list on `stable-diffusion.cpp` (which I intend to use as the inference engine for this), I noticed that compared to LLMs - where a pass of Heretic will remove refusals and thus "uncensor" them - Diffusers don't really seem to have that... Granted, they also consist of many more different parts - VAE, Encoder, ...

So, I want to understand this: Without using a LoRA, which models are uncensored by default?

Because of course, a LoRA can "fix" a model by adapting it, obviously. But to be honest, I have not really figured out all the moving parts just yet. Sure, I see them in ComfyUI workflows - but due to my visual impairment, those spaghetti graphs just absolute _eliminate_ me. xD Like, literally, headache. So I hope that by using a purely simple tool like sd.cpp and use it as a tool in LibreChat, I can replicate the generation flow of ChatGPT/QwenAI (their official chat.qwen.ai frontend, idk what to reference it as lol).

I still have a lot to learn in terms of Diffusers; so any good pointers to read about the various parts and knobs would be nice along the way. :)

Thank you!

document fetched 2026-10-20 00:31 · canonical https://reddit.com/r/StableDiffusion/comments/1xa79n4/does_a_purely_uncensored_model_exist_without_a