Launching today
lucidml
World models that run on your own GPU
1 follower
World models that run on your own GPU
1 follower
Move through an open world that a neural network generates frame by frame as you press keys. There is no game engine underneath. I trained it from scratch on my own architecture, Summaryformer, where every frame writes its own memory. Play it in the browser, or download the open weights (Apache-2.0, 630M parameters) from Hugging Face and run it on your own NVIDIA GPU!
Free
Launch Team
Render Sandboxes: a computer for every agentIsolated compute, persistent state and a real filesystem
Promoted
Hi, I'm Abhishek. I built and trained this model on my own.
Play it here, nothing to install: https://lucidml.ai/dashboard/minidemo?world=gta5
It's a 630M-parameter foundation model that I trained from scratch. It was pretrained on 100 million images, 4 million videos and 200,000 very long videos, and then trained on about 800 hours of GTA V gameplay (100,000 videos) with recorded controls. Nothing is fine-tuned from someone else's model.
You press keys and it generates the next frames at 384×672. There is no game engine underneath.
It runs on a new architecture I made called Summaryformer. Every frame writes 96 summary tokens about itself. The last 2 frames stay in memory whole, and older frames are kept only as their summaries. I wrote down how it works here: https://lucidml.ai/summaryformer
The weights and the inference code are open under Apache-2.0: https://huggingface.co/lucidml/gta. Running it locally needs an NVIDIA GPU.
I learned how KV caching works for video from the Self-Forcing repo, so a big thanks to them.
If you run it, reply with your GPU and frame rate. I'd also like to hear what you would want to do with a model like this.