World Labs Open-Sources Marble Generative World Model and Introduces Chisel for 3D World Editing

World Labs has released Marble, an open-source generative world model that simulates dynamic 3D environments, alongside Chisel, a powerful editing tool for customizing these worlds. Ideal for robotics, AR/VR, and beyond.

Discovering Marble: A Breakthrough in Generative World Models

World Labs, a pioneering AI company co-founded by renowned computer vision expert Fei-Fei Li, has made significant strides in spatial intelligence. Recently, they open-sourced Marble, a cutting-edge generative world model designed to create and predict behaviors in expansive 3D environments. This release marks a pivotal moment for researchers and developers working on embodied AI, where understanding physical spaces dynamically is crucial. Alongside Marble, they've introduced Chisel, an intuitive editing tool that empowers users to modify these simulated worlds with precision.

For beginners, think of Marble as a 'digital twin' creator. It doesn't just generate static images; it simulates how a 3D world evolves over time based on real-world actions, like moving a camera or commanding a robot. This is foundational for applications where AI needs to anticipate changes in complex, real-world settings.

The Core Capabilities of Marble

Marble excels at predicting future states in large-scale 3D scenes. Given an input of RGB images paired with camera poses, it generates subsequent frames that reflect realistic environmental dynamics. For instance:

• Camera Movements: Simulate panning, zooming, or rotating through a cityscape, with Marble forecasting photorealistic next frames.
• Agent Interactions: Predict outcomes from robot actions, such as navigating obstacles or picking up objects.

This predictive power stems from its training on a massive dataset: high-fidelity 3D reconstructions derived from scans of numerous U.S. cities. These scans capture intricate street-level details—pedestrians, vehicles, buildings—providing a rich foundation for learning spatial and temporal patterns.

In practical terms, imagine training a delivery robot. Instead of relying solely on real-world trials (which are costly and risky), developers can use Marble to simulate thousands of scenarios rapidly, iterating on behaviors in a safe, virtual space.

Diving Deeper: Architecture and Training

From an advanced perspective, Marble represents a video diffusion model tailored for world simulation. It processes sequences of images and poses to output future video frames autoregressively—one frame at a time, building coherent long-term predictions.

Key technical highlights:

• Input Format: Multi-view RGB images + 6DoF camera poses (position and orientation).
• Output: High-resolution future frames that maintain geometric consistency and physical plausibility.
• Dataset Scale: Millions of 3D scans processed into structured reconstructions, enabling generalization across diverse urban environments.

Training involved optimizing for long-horizon predictions, which is challenging due to error accumulation in sequential generation. Marble mitigates this through advanced diffusion techniques, ensuring stable rollouts over dozens of steps.

You can explore the model firsthand via its GitHub repository: worldlabs/marble. The repo includes pretrained weights, inference code, and scripts to replicate experiments.

Benchmarking Marble's Performance

To validate its prowess, World Labs evaluated Marble against established baselines like VideoPoet and LWM. Results on custom benchmarks for urban navigation and object interaction show Marble outperforming competitors:

These metrics highlight Marble's edge in fidelity and consistency, making it suitable for precision-demanding tasks like autonomous driving simulations.

Real-world example: In a demo video, Marble simulates a drone flying through a bustling street. As the drone 'banks left,' the model generates frames showing updated building facades, moving cars, and lighting changes—all without predefined maps.

Introducing Chisel: Precision Editing for Simulated Worlds

Complementing Marble is Chisel, a suite of tools for interactive world editing. Chisel allows users to sculpt generated environments post-generation, addressing a key limitation in pure generative models: lack of fine-grained control.

Chisel's features include:

• Object Removal/Insertion: Erase a parked car or add virtual pedestrians seamlessly.
• Style Transfer: Transform a daytime scene to sunset or apply artistic filters while preserving 3D structure.
• Semantic Edits: Modify high-level attributes like 'make the street busier' via natural language prompts.

Under the hood, Chisel leverages segmentation masks, inpainting diffusion models, and 3D-aware lifts to ensure edits propagate consistently across views and time.

For developers, getting started is straightforward. Clone the repo at worldlabs/chisel, install dependencies (PyTorch, Diffusers library), and run:

git clone https://github.com/worldlabs/chisel.git
cd chisel
pip install -r requirements.txt
python demo.py --input_path marble_output.mp4 --edit 'remove red car'

This generates an edited video in seconds. Advanced users can extend it with custom LoRAs for domain-specific edits, like industrial sites or indoor spaces.

Real-World Applications and Broader Impact

Marble and Chisel unlock transformative use cases:

• Robotics: Accelerate sim-to-real transfer by generating diverse training environments.
• AR/VR: Create infinite, interactive worlds for immersive experiences.
• Autonomous Systems: Test edge cases in simulated cities without hardware.
• Gaming/Entertainment: Procedurally generate dynamic levels with editable elements.

Fei-Fei Li's vision emphasizes 'spatial intelligence'—AI that truly understands 3D spaces like humans do. By open-sourcing these tools, World Labs democratizes access, fostering innovation across academia and industry.

Consider a practical workflow for robotics devs:

1. Generate base world with Marble.
2. Edit via Chisel for specific scenarios.
3. Train policies in the sim.
4. Fine-tune on real data.

This loop reduces development time from months to weeks.

Challenges and Future Directions

While impressive, Marble has limitations: It's optimized for urban outdoor scenes, with potential for indoor or off-road extensions. Computational demands are high (A100 GPUs recommended for inference). Future work may include multi-agent simulations or integration with reinforcement learning frameworks like MuJoCo.

Community contributions are encouraged—check the GitHub issues for ongoing discussions on quantization, faster inference, or new datasets.

Hands-On: Building Your First Marble Simulation

To make this actionable, here's a beginner tutorial:

1. Setup Environment:

   conda create -n marble python=3.10
   conda activate marble
   git clone https://github.com/worldlabs/marble.git
   cd marble
   pip install -e .
   

2. Run Inference:

   from marble.model import MarbleSimulator
   simulator = MarbleSimulator.from_pretrained('worldlabs/marble-base')
   frames = simulator.predict(
       rgb_sequence=[img1, img2],
       poses=[pose1, pose2],
       action='move_forward_2m',
       num_steps=10
   )
   

3. Edit with Chisel: Integrate outputs directly for post-processing.

Demos and videos are available in the repos, showcasing everything from city flyovers to robot pick-and-place.

In summary, Marble and Chisel represent a leap in generative world modeling, blending prediction, simulation, and editing into a cohesive toolkit. Whether you're a researcher probing AI limits or a practitioner building the next robot fleet, these open-source resources provide a solid starting point for innovation.