wqslave wrote: ↑Wed Jun 29, 2022 4:55 pm
Beta.exe's approach is infinitely configurable, it just also works with a noun/adjective system, also with categories. So it looks at everything you've provided for the nouns/adjectives/names/ect, and each base sentence, and creates captions with regex replacements. So to make a caption there, you create a template sentence with spots for a bunch of parts that are put together at runtime. To do the same here you'd need to input each potential combination individually, which can quickly add-up to several thousand lines. It's not like a huge issue or anything that makes this unusable or something, it just isn't as robust.
Highlighting is an alternative thing that beta.exe and some other censor-related softwares do where they put a box [] or circle () ect around a part. Kind of like a very basic identification box, but without a label. Beta.exe utilizes multiple censor styles (blur/black box/pixel) and then also highlight. It's additional sensory overload and/or to highlight "where you should be looking."
Oh, I see. I didn't actually mess with beta.exe for very long. I wasn't aware you could do any manipulation of the captions. I had thought about implementing a system along the same lines, but my inspiration for that was actually the Vocabulary system from good ol' TeaseAI. If there is enough interest, I might do that. Priority right now is bugfixes. When I have time to actually do an update followed by proper testing, hopefully those will be all worked out, or at least the major ones.
My dream implementation for the captions would be to train a NLP model on a few million bdsmlr/tumblr captions (which would require an OCR model to do all the text extraction...
if you give a mouse a muffin... ) and then have the program hit an API for a caption generated by a virtual humiliatrix... or whatever the user is into. But I'll be honest, as much as I would like to set it up, it's probably not going to happen due to time and expense involved. Although, I could think of a few dozen other uses for the same NLP model, and I do have sufficient infrastructure at my disposal...
The highlighting is probably fairly simple to implement, now that I think about it. The Nudenet library is just using PIL to draw some boxes over the coordinates it returns as matching a selected tag. But it could draw anything, really. The difference is just switching up a couple of function calls. I like low-hanging fruit. Let me look into this one this evening when I'm done with actual work and have time for pet projects.