CivArchive
    Qwen3 VL Instruct 4b Heretic 7refusal +convrot (alternate Krea 2 text encoder for ComfyUI) - v1.0
    NSFW
    Preview 134765790
    Preview 134775092

    An abliterated text encoder for Krea 2 or any other model that uses Qwen 3 VL Instruct 4b.

    Note that this is neither necessary nor sufficient to make NSFW generations, but it may be better at expanding NSFW prompts, and I've preferred results from abliterated models before. (After experimenting, it seems like it may make your results slightly more nsfw if you're otherwise not doing any other jailbreaks, but not reliably.)

    Not guaranteed to be better than the normal model. I'm just uploading it because I did the conversion and thought I'd share in case anyone else needs it.

    HF link to original model in diffusers format: https://huggingface.co/svjack/Qwen3-VL-4B-Instruct-heretic-7refusal

    Description

    FAQ

    Comments (12)

    azra1lJun 24, 2026
    CivitAI

    so this model works with the generate text node inside comfyui?

    _Envy_
    Author
    Jun 24, 2026· 3 reactions

    Yup. It also works as a replacement text encoder for Krea 2.

    azra1lJun 24, 2026

    sweet baby jesus!

    SD_STORMJun 25, 2026
    CivitAI

    Please add more comparison images with normal\abliterated

    _Envy_
    Author
    Jun 25, 2026

    In the middle of some training, but I'll add some in a bit.

    ArchAngelAriesJun 25, 2026
    CivitAI

    THANK YOU!!! ❤️❤️❤️

    velantegJun 25, 2026· 9 reactions
    CivitAI

    testing shows default TE gives more quality and actually not censoring anything.

    _Envy_
    Author
    Jun 26, 2026· 1 reaction

    I mentioned in the description that this isn't necessary for NSFW. I don't see any quality issues with it, so you're welcome to post your own comparisons.

    GPUPoorChadJun 28, 2026

    Only real use is if you want to use the default prompt enhancer without it bitching an don't have insane amounts of VRAM, the quality loss isn't that much. But yeah, I don't think text encoders can refuse currently, just shifting the output slightly to something the diffusion model expects slightly less

    redhomeJul 5, 2026

    I actually tried using the Qwen 3 VL 8b (Heretic) and it gives much more tamer results than using the 4b scaled version, even when prompted not to. Really strange.

    _Envy_
    Author
    Jul 6, 2026

    @redhome Did you mean 4B? You shouldn't be able to just swap in 8B.

    redhomeJul 8, 2026

    @_Envy_ Actually, the 8b.