Lucid

We are building universe simulations powered by interactive video models. We train video models that simulate hyper-realistic environments with immersive control, replacing hard-coded game or physics engines with dynamic neural networks. We built the fastest action-conditioned diffusion video model (running at 20+fps on a 4090 gaming gpu) to simulate minecraft. It is 5x faster than other minecraft World Models and was trained with 100x less resources. Our unique insight was relying on aggressive compression in our tokenizer (128x versus the traditional 8x), and because attention scales quadratically with # of tokens our model can run blindingly faster. Now we’re training a hyper realistic world model!