lucidml

Realistic World Simulation on Deep Neural Networks @NVIDIAAI

Throwback to the first model I made in December. Even though I had made it then, I only tested it recently In February. 100% Realtime on a RTX 5090. This is one of the oldest iterations of the world model before we started training on multiple games datasets. It's funny how the model keeps hallucinating random things and comes back to the GTA manifold after a while lol, its a bit like LSD (I think?) What I liked a lot here was that it was the first proof of concept I ever had. I was moving in a world completely simulated by a neural network, trained on Google Cloud Credits. Temporal context was 30 past frames. I kept playing it, got a bit lightheaded and then the Rope position embedding reached it's limit after 2 minutes, that's when I finally stopped. I was so happy. @lilshake139_ #lucidml #worldmodels

@_SRI_VISHNU Yes! We definitely can, once we raise our next round of funding we'd love to work on this.

@lucidmlx Image models made in other countries do not get indian context they don't know who arjuna or dasharath is ,can you train a image model more on indian data

In March, we asked a simple question Can we train an image model from scratch at a fraction of frontier models' spend? Today, we’re releasing our first answer. We trained our Medium Image model for just $4000. No distillation. No frontier model shortcuts. No massive cluster. Just clean training + careful data + ruthless efficiency. For context: MosaicML spent ~$48,000 to replicate Stable Diffusion 2. That gap shouldn’t exist. This is our first step. World models and video are next. All images below are raw outputs from the model. Try it → lucidml.ai/image Tech report → lucidml.ai/imagetechreport

This is just the beginning ! In the next few days we'll release an image model and after that an endpoint to stream and try out world model too! Thank you @caleb_friesen