Even if you abliterate your model using the old abliteration script or the newer heretic, I found that the models still feel somewhat censored as they purposefully avoid using specific styles and vocabulary, as if Deepmind/Qwen et al have entirely stripped or replaced "bad" words or texts from their corpus of training data.
A related blog post (https://news.ycombinator.com/item?id=47842021) discussed this and termed it "flinching". I wonder if this flinching could also be "mediated by a single direction" or if it can only be fixed by finetuning on a more extensive text corpus.
Unfortunately, Glaze does not seem to work. When I've trained a simple style LoRA on a few sets of glazed images using SDXL, the LoRA was still able to reproduce their style.
Another unfortunate consequence of the introduction of Glaze and Nightshade is that some artists which I follow have now started glazing all of their new works which they publish, leading to quite ugly results due to the noise that Glaze produces on high settings, despite questionable efficacy.
A related blog post (https://news.ycombinator.com/item?id=47842021) discussed this and termed it "flinching". I wonder if this flinching could also be "mediated by a single direction" or if it can only be fixed by finetuning on a more extensive text corpus.