@antopatrex1 I thought that the folks building the 3d environment setups were solving this problem to varying degrees? Damn no r2d2 for a while I guess
Iβve been thinking of the possibility of only inserting understanding into a model while ripping out the underlying data layer, so we are able to build out the scaffolding for logic while maintaining small compute requirements, would be cool to think about how to achieve this, and open to discussion ofc (mostly because this sounds kinda out of whack and I donβt want to be the crazy guy)
I am also building image gen for 3d gauss splat on Prem and condensing it to make it significantly better than depth anything v3, will I fail? Letβs find out