/vt/ - Virtual Youtubers

>>83921463
>(if you want a very high quality model)
That's the point, kinda.
>but it could be very slow
I suspect that this would not be slow at all once you give Unet a hint somehow. It already knows the characters, but old senile boomer pony's unet just doesn't understand alpha sigma slang. Maybe a dreambooth-style finetune for a few epochs is the key here after all.
>No?
>Adding extra noise breaks zeroSNR. Model learns to expect and remove this extra noise.
I mean, no - yes? It doesn't add the noise to anything but the "input" latents, that's what "I" is for? Why should that cause any issues with ZSNR? If anything, it should just learn to produce slightly noisy image every time, but due to the fact that the added noise also changes every time it just doesn't do that. And well, it learned to do solid black/white backgrounds perfectly fine (still struggles with magenta though).
>optimi
I think I might have heard of it on /lmg/ or /hdg/ once but this time I found it while searching for 8bit stable optimizers, not sure if it even works for sd-scripts