Small refinement: the underlying model isn’t stochastic at all. The forward pass is a deterministic function of the weights and input, it just produces a probability distribution over the next token. The stochasticity is an optional sampling step layered on top, not something inherent to LLMs. Greedy/argmax decoding (or temperature 0) makes the whole thing deterministic.
So “purely stochastic” overstates it a bit: the distribution is computed deterministically, and you choose whether to sample from it or not.
Maybe dumb question but how do I just take a sat image and create the scene? The scripts in the repo are all about training which I assume requires you to have the 3d data too.
These sort of projects always look cool but I think the real "wow factor" would be a file upload where you can see the result on your image. I assume there are reasons why this isn't done.
Is there anyone monitoring US government websites for deletion of pages beyond what people may notice in an ad hoc way? I feel like this should be automated.
So “purely stochastic” overstates it a bit: the distribution is computed deterministically, and you choose whether to sample from it or not.