"Unbounded video games and fully differentiable reinforcement learning gyms."
TL;DR: Lucid is building generative simulations powered by fast world models. Instead of using traditional game engines with hard-coded physics, their models learn to simulate reality from pixels, enabling real-time interactive environments. With it they will train robots in their own imaginations and make unbounded gaming experiences. They trained the fastest world model ever seen to simulate minecraft end-to-end (20+fps on a gaming GPU).
The Problem: Game Worlds Are Static & Expensive to Build
Modern game development is slow, expensive, and constrained:
GTA V took 3 years, 1,000 employees, and $100M+ to build—AAA game budgets are skyrocketing and they’re not getting any better.
Despite the price tag, these games are inherently static, with predefined environments, objects, and interactions.
Players can’t truly shape the world—every door, street, and event is pre-scripted.
Meanwhile, robotics faces its own bottleneck—AI models trained in simulators (MuJoCo, Isaac Sim, Gazebo) fail to generalize to the real world (Sim2Real gap) because today’s simulations are hand-coded approximations of physics rather than learned from real-world data.
Their Solution: Generative World Models
Lucid replaces traditional game engines with a generative simulation engine that learns from data rather than being manually programmed.
Every frame is generated in real-time, conditioned on player actions.
Trained on video, not game scripts—their models learn the rules of physics directly from pixels rather than hardcoded logic.
Infinite, dynamic game worlds—players can generate and explore entirely new environments just from a text prompt or sample concept art.
A Neural Minecraft Simulator
They trained a neural network to simulate Minecraft end-to-end—every pixel is generated in real-time, learned from 200 hours of gameplay.
Runs at 20+ FPS on an NVIDIA 4090—5× faster than existing world models (Decart’s Oasis <4 FPS).
Aggressive latent compression—they utilize a VAE with 128x spatial compression allowing them to vastly reduce the amount of tokens needed to represent a single frame
What’s Next? Training on the Real World
They are now training their models on real-world video data to build a general-purpose universe simulator for:
Gaming: The last game engine humanity ever needs—generating unique environments dynamically from simple text or multimodal prompts.
Robotics: Simulations that actually match reality—training embodied AI models in diverse, realistic environments. A fully differentiable, learned simulation framework for reinforcement learning.
💥 Are you working on AI/robotics and need high-fidelity simulations? Lucid is selecting early partners to fine-tune LoRAs on domain-specific data. Want to explore the future of generative gaming? Sign up for early access to Lucid v2.
🤝 Interested and want to connect? Reach out to the founders here.
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
.png)
