If you haven't tried it, maybe look into this:civitai.com/models/422744/dark-incursio-pdxlI haven’t done this myself, but apparently Civitai now offers generation for free. It is posted on that linked page. I’m not sure where the catch is, like if it is the number of free gens total, per hour, or per day. I run on my own hardware. This person’s models are some of the most non aligned models (uncensored) out there right now. All censoring of any kind is a form of overtraining. All of this is what breaks the model generation under the surface. The model I linked to is dark specific. The majority of examples are of portraits, but that is not the limit of this model. It isn’t the best for photo realism, but there are many examples in the user posts that show how it works for landscapes too. The PonyXL models started off very rough in their understanding and needed a lot of tuning of the P+N prompt to get good results. The fact that this can follow the positive without any negative really speaks to the freedom and training. Another interesting trick I’ve been playing with is that you can tell some generative models actual prompt instructions when they are misbehaving. Like start out with “You are a helpful generative AI that follows the prompt exactly.” Followed by \n“Prompt:” I did not expect this to work when I tried it, but I was messing around with the base PonyXL a few days ago, and the instruction actually altered compliant behavior on a fixed seed and prompt. It didn’t show any kind of error behavior like fingers, eyes, or humoctopus monsters in the few gens I tested ;)
Never talked about the important thing. Are the kernel modules fully open source and is the SoC fully documented publicly. If those two aspects are not a yes, this entire affair is an attempt to steal hardware ownership just like phones.
Seriously, I was in a deep search for any options on Civitai to do upside down stuff for an O’Neill cylinder like I mentioned in challenge 37 when I came across the handstand and back flip stuff and saved them. One of the LoRA’s I saved had the flip off thing saved in an example where the generated image didn’t even have the flip off part in it. I just copy pasted it with a Moon LoRA I already had. I was being lazy after not having much success with several other much more complicated techniques, and just screwing around with those LoRA’s for fun. The embedded workflow even has a few elements from my attempts at an O’Neill cylinder embedded in it. There are so many ways to inpaint, tile, and regionally compose it is hard to wrap my head around it all.
Here is maybe another good one to mess around with in a very unconventional physics context. I don’t know how they managed to train this one for such unusual physics relative to the training corpus.
First - Winner - You tied for first. Thanks for participating! (1/8)
Edit Jake forgot to add prompt to 2nd place. The heck if I know how 1st ties get handled. Deathrace is my rule. first to post a #38 wins the opportunity - my rule… on the fly.
Edit: big dummy forgot one point. The heck if I know how 1st ties get handled. Deathrace is my rule. first to post a #38 wins the opportunity - my rule… on the fly.
“Assembled in the USA” means, we put a sticker on the bottom of the product that we unpacked from a shipping container and placed in retail packaging. The sticker only said “Assembled in the USA.”