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.
That's likely not a trained behavior, though, it's probably the result of filtering the training data. It's not "when these parameters fire, trigger a refusal", it's the absence of parameters triggering the flinched words in the first place.
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.
If OpenAI steals all your work, that's copyright infringement - but if you tried to stop them through technical means and they do it anyway, that's felony DRM circumvention.
I was a better person this morning for not knowing that furries had the term "hindquarters". I mean, that's fine for other people, you do you, but for me, I was better this morning.
The v1.3 model weighs in at 4.3 GB. There's an additional download of 1.6 GB of other models due to usage of huggingface's transformers (only once on startup). And the conda env takes another 6 GBs due to pytorch and cuda.
Larger images will require (much) more than 5.1 GB. In my case, a target resolution of 768x384 (landscape) with a batch size of 1 will max out my 12GB card, an RTX3080Ti.
I think this is a good time to ask if anyone is working on parallelizing machine learning compute anymore? For at-home computation like this it seems like it would be a lot better to allow people to stack a few cheaper GPUs rather than having to pony up thousands of dollars for ML-oriented beast cards to be able to do things like generate large images.
For videos in particular, if you don't mind shelling out cash, the current go-to (at least according to various AI discord servers I'm on) for AI animation nerds is currently Topaz upscaler. There are free alternatives but I've yet to see any of them work as well as Topaz, though I'm sure that will change soon. For interpolating frames Flowframes is "free" (new features if you join the Patreon) and is IMO very good.
I've seen a number of 80s/90s VHS recordings of concerts being uploaded to YouTube in 4K (using Topaz) and they look like they were recorded that way, truly amazing. I do hear it can be a bit of work though getting the settings right.
I frequently hear that screen protectors which mimic paper surface lead to a much faster abrasion of the tip of the Apple pencil. Is this the same in your case?
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.