Post #1468607

@Njord I wish there was an ethically trained, locally usable diffusion image model. A couple of years ago I was very much into using SD to generate images to use in TTRPG, mainly character portraits. It was nice that I could do this locally, and using no more power than if I fired up a video game on my PC. In reflecting on that today, I think what was actually useful about it may have been the effort of writing the prompts. Seeing the image served as some kind of confirmation that what I was seeing in my mind's eye was a lot like what others at my table would imagine, even if I didn't share the image. Things that made it just a toy include that you couldn't do things like "now put character A and character B into setting X" and get recognizably the same people as in their original portraits. Heck, you couldn't even get the same portrait twice. Too much randomness. An LLM-adjacent tech I wonder about is cosine-similarity search. Would I benefit from this as a complicated alternative to full-text search for local code searching? Finally, I have found ente.io's local image classification ("magic search") to be quite satisfactory, once I accept that it's just a best-effort search. It makes all kinds of errors ("dog" gets you all sorts of non-dog images) and there's no way of knowing if it's comprehensive (these can't really be ALL my "macro" photos here, can they?) but it does let me quickly find relevant photos from my large collection using obvious terms. I don't know anything about the origin or training of this model and I haven't thought deeply about how flaws in the model would have real world impact. But for instance if you told me that this model turned out to have a racial bias when searching for occupation names I would not be at all surprised...