Or maybe even every session?
Fuck that… every frame.
Although, that was the experience with Stable Diffusion a few years ago. WAN Video is good enough for remembering the face during the course of the processed length (4-5 seconds or so). But, as soon as it needs any sort of object permanence, like when the face is hidden, suddenly, it doesn’t know what it was looking at any more. Seems like a solvable problem with reference images. I doubt NVIDIA thought that far, though.
sobchak@programming.dev 2 days ago
NNs are deterministic. Chatbot and image generator implementations just purposely add randomization to make them seem more intelligent.
Buddahriffic@lemmy.world 2 days ago
Though depending on how much context they pull in from neighbours, you can get overfitting, like a character might end up with a different face depending on what’s in the background. Vague unreadable text could be replaced with random readable text that also heavily depends on what is seen elsewhere in the frame (assuming it’s trained on enough text that if doesn’t just make legible text illegible).
So while it can be deterministic, it can still be highly chaotic, depending on how exactly it gets implemented. And I’d guess that it’s a double edged sword where sometimes more context gives better results and other times less context gives better results. And I’d also guess that if you want to avoid overfitting, the amount of training data needs to increase exponentially with the number of inputs, assuming there’s enough neurons in the network to encode the necessary complexity.