santa hat
deerdeer nosedeer glow
Sign In

Before You Downvote Me for Creating Checkpoints in the Age of Textual Embeddings


EDIT: Okay, enough. I carved out some time to install Kohya_ss, and am in the process of converting the checkpoints to LORAs. So quit with the aggressive comments. It would have been nice if some of the more unpleasant and really self-entitled commenters had actually offered some help instead of just flipping the bird and hitting the down-vote button. But in the end, those guys live in a self-punishing state, so I guess I'll just let auto-karma handle it.

Guys, I really admire the work of people like FranklyFrank2K. I think it's great these things can be achieved in 30k, or 10mb etc. It's miraculous.

Unfortunately I learned DreamBooth workflow when I had more time in my life, and before Textual Inversion was any damned good, and when LORA was just coming into being.

I know that it would take a good long evening for me to learn how to make a basic textual embedding, maybe another to learn LORA.

Then it would take some weeks or even months to get any good at either/both of these, which have a text/semantic component that's not so evident in DreamBooth.

Guys- I am not proud of being as time-poor as I am, but I don't even have that initial evening to spare any more. There's real life, and holidays, and people I want to be with in the real world, and I have to work, and sometimes I have to sleep.

Plus all the tutes seem to be really long YouTube videos, instead of step-by-step SubStack posts, which would speed the process up. I get why - likes, engagement, influence, and more sponsor potential. But anyway, that's an additional barrier for me.

So to those angry folks who down-vote me for doing under-represented celebs in a heavy format like a safetensors checkpoint - I'm not doing it to torment you. I just honestly don't have time to learn this wheel again. What I'm putting up in this period is 'archival' work. I hardly have time to mess about with SD anymore.

And I do encourage anyone who knows how to do it, to convert any of my models into a lighter format, if you know how to do it, and upload it. If you do, I'd be glad if you let me know, I'd love to see it.


PS Once SD XL is bedded in, I expect that any future celebrity character projects I could find time for will be LORA, or whatever the semantic, non-checkpoint flavor of the day will prove to be, once the dust has settled.

PPS - On advice of a commenter here, I do intend to use the next free evening I have (heaven knows when that will be) to install Kohya SS, after which I will convert these checkpoints to LORAs and upload them if they are satisfactory. But free evenings are hard to come by!