Uncensored models via LORA
Then for Krea2 turbo/base I saw that there are LORAs to remove censorship in the text encoder - well, image models is not something that is often discussed here, so I am not surprised to not have read about that here.
Proof1 community thread
Key points
- I know the heretic tool is the usual generic approach to remove refusals from a model, and a lot of research has been done to find out what parts in the LLM need to change.
- This gets very expensive though the larger the model is.
- It seems just to reduce refusals for cyber purposes, but I think the LORA approach makes it cheaper to create and more flexible (just load/unload a lora, keep the model as-is)?
- Has anyone done experiments between this for eg glm 5.3 Flash and the heretic or cyber model versions of it?
Sources (1)
- [1]Uncensored models via LORAr/LocalLLaMA (top, daily) · Oct 17, 06:57 AM
Then for Krea2 turbo/base I saw that there are LORAs to remove censorship in the text encoder - well, image models is not something that is often discussed here, so I am not surprised to not have read about that here.
I know the heretic tool is the usual generic approach to remove refusals from a model, and a lot of research has been done to find out what parts in the LLM need to change.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 17, 2026Show HN: Llmspeed – feel how fast a local LLM runs before you buy the hardware
- Oct 16, 2026Comfy-Org/MOSS-Image
- Oct 15, 2026open source framework that makes your AI agents generate UI elements
- Oct 15, 2026Granite Switch 4.1: Chat, Decisions, and Speech from One Endpoint
- Oct 15, 2026NVIDIA Nemotron 3.5 Super VL Powers Specialized Agents with Multimodal Reasoning
- Oct 15, 2026I put a conformal risk control gate between an LLM agent and its write actions on tau-bench. The bound held. The gate didn't beat asking every time [R]