Embed Notice

@GossiTheDog What's striking is how obvious it is that they're doing this entirely in the wrong domain, purely as images. A viable model would correlate the training material to rigged 3d model, learn the motions in this constructed 3d space, do generative stuff in this space, and then use a mix of the model and non-AI rendering techniques to carry it back to 2d video.

The reason they don't do this is that they're high on their own supply and believe the deep learning bs will do all that itself