Runway's GWM-1 Models: Generating Long Videos with Realistic Physics for Robotics and Entertainment

Runway unveils GWM-1, advanced world models that produce minute-long videos with consistent physics, powering robotics simulations and cinematic content creation.

The Dawn of Physics-Aware Video Generation

Imagine crafting a video where a robot arm deftly stacks colorful blocks without defying gravity, or a cartoon character bounces realistically across a trampoline for a full minute. Traditional video generation tools often falter here, producing clips with erratic motion or crumbling physics after just a few seconds. Enter Runway's latest innovation: General World Models (GWM-1). These models mark a pivotal shift in AI-driven video synthesis, enabling the creation of extended sequences that obey the laws of physics consistently.

World models represent a sophisticated class of AI systems trained to predict future states of dynamic environments purely from visual data. Unlike standard diffusion models that generate frames independently, world models learn latent representations of physical interactions, allowing them to simulate plausible continuations over time. Runway's GWM-1 builds on this foundation, trained on vast datasets to produce high-fidelity videos up to 60 seconds long— a significant leap from the typical 5-10 second limits of prior tools.

Unveiling the GWM-1 Family

Runway has released two variants to cater to different needs:

• GWM-1 Preview: The flagship model, delivering top-tier quality for professional applications. It generates videos at 480p resolution (854x480 pixels) with 17 frames per second, maintaining coherent physics throughout.
• GWM-1 Mini: Optimized for speed, this lighter version runs inference up to 10x faster on consumer hardware, ideal for rapid prototyping or real-time uses.

Both models support flexible conditioning: start with text prompts, images, or short video clips to guide generation. For instance, prompting "a red robotic arm stacking blue and yellow blocks on a table" yields smooth, physically plausible animations where objects collide, roll, and balance naturally.

To access these models, developers can dive into the official repository at GitHub, which includes code, pretrained weights, and inference scripts. This open-source release democratizes access, allowing researchers to fine-tune or extend the models for custom domains.

Training on a Massive Scale

The power of GWM-1 stems from its training regimen. Runway curated over 40 million video clips, spanning diverse sources:

• Everyday internet videos capturing real-world physics (e.g., sports, accidents, object manipulations).
• Specialized robotics datasets from platforms like DROID and ROBOTURK, featuring dexterous manipulation tasks.
• Synthetic simulations to augment rare events like complex multi-object interactions.

The architecture employs a video latent diffusion model with a custom tokenizer that compresses 14-frame chunks into compact representations. During training, the model autoregressively predicts subsequent chunks, enforcing temporal consistency. Key innovations include:

• Physics regularization: Losses that penalize violations like objects passing through surfaces.
• Long-horizon forecasting: Techniques to propagate dynamics over hundreds of frames without drift.

This results in emergent capabilities, such as accurate trajectory prediction for thrown balls or stable cloth draping.

Rigorous Evaluation and Benchmarks

Runway didn't just claim superiority—they backed it with benchmarks. On physics-focused tests:

• PhyScene: GWM-1 outperforms baselines like Stable Video Diffusion by 2x in trajectory accuracy for rigid bodies.
• ParPhyBench: Excels in parallel physics scenarios, like multiple bouncing balls maintaining independent paths.

For robotics, evaluations used simulated tasks:

• Block Stacking: Generates sequences where a robot successfully builds towers 80% more reliably than competitors.
• Object Grasping: Predicts grasp success with 15% higher fidelity.

Qualitative demos showcase versatility:

Example Prompt: "A humanoid robot jumps repeatedly on a trampoline in a gym."
Output: 60s clip with elastic bounces, arm swings for balance, and no mid-air glitches.

Another: "A squirrel navigates a parkour course with logs and branches." The model simulates grip friction and momentum shifts flawlessly.

Real-World Applications: From Robots to Reels

Robotics Revolution

In robotics, accurate simulation is gold. Traditional physics engines like MuJoCo require manual parameter tuning and falter on novel objects. GWM-1 offers data-driven alternatives:

• Policy Learning: Train RL agents in generated videos, transferring to real hardware with minimal sim-to-real gap.
• Prospective Planning: Simulate robot actions forward in time to select optimal trajectories.

For example, researchers could input a live camera feed of a robotic arm, append a desired goal image, and generate preview videos of potential maneuvers—accelerating iteration cycles.

Entertainment and Content Creation

Filmmakers and animators gain a new tool for previsualization:

• Consistent Characters: Generate extended shots where actors maintain gait, clothing folds, and expressions.
• VFX Prototyping: Mock up destruction sequences or crowd simulations with realistic debris and collisions.

Practical workflow:

1. Sketch a storyboard image.
2. Prompt GWM-1 with text like "camera follows a cyclist down a winding mountain road at dusk."
3. Refine with inpainting for specific edits.
4. Export for final production.

This bridges the gap between static AI art generators and full CGI pipelines, slashing costs for indie creators.

Looking Ahead: Scaling the Simulation Frontier

GWM-1 is just the preview (pun intended). Runway outlines ambitious roadmaps:

• Longer Durations: Push to 5+ minutes for narrative storytelling.
• Higher Fidelity: 1080p+ with audio integration.
• Controllability: Direct manipulation of object properties (e.g., "make the ball bouncier") via editable latents.
• Multi-Modal: Incorporate depth sensors or robot telemetry for hybrid real-virtual sims.

Challenges remain, like handling occlusions or adversarial lighting, but the trajectory is clear: world models will underpin the next era of embodied AI.

Get Started Today

Ready to experiment? Install via pip from the GWM-1 GitHub repo:

git clone https://github.com/runwayml/gwm1
git checkout preview  # or mini
pip install -e .

# Quick inference
python scripts/inference.py \\
  --prompt "robot stacking blocks" \\
  --output video.mp4

Join the community on Hugging Face for model cards and discussions. Whether you're a robotics engineer debugging grasps or a YouTuber crafting viral clips, GWM-1 equips you with unprecedented creative and analytical power.

This isn't merely video generation—it's virtual physics at your fingertips, unlocking simulations that feel indistinguishably real.