Runway has achieved a technical milestone that changes how AI video systems work: moving from offline generation that takes 30 to 60 seconds to real-time video creation that responds instantly during live conversation. The company released Characters, an avatar experience built on its new general world model (GWM-1), which generates video frame-by-frame at 24 frames per second with latency under 160 milliseconds. This represents a fundamental shift in how AI video models are architected and deployed.

What Makes Real-Time Video Generation So Difficult?

Until now, every major AI video model, including Runway's own Gen 4.5, operated offline. These systems generate an entire video sequence at once using what engineers call "bidirectional diffusion," meaning the model has access to context from both before and after each frame during generation. This approach works well for pre-recorded content but makes real-time interaction impossible because the model must wait to see the entire sequence before producing any output.

Real-time video generation requires a completely different architecture. The model must predict frames one at a time, using only information from previous frames, without knowing what comes next. This is called "causal" generation, and it's significantly harder to train than offline models. During training, causal models can only look backward in time, not forward, which removes a crucial source of context that makes frame prediction easier.

How Runway Built the First Autoregressive Video Model

Runway's solution was to develop GWM-1, an autoregressive video model that works similarly to how large language models (LLMs) generate text. Instead of generating all frames at once, GWM-1 predicts the next frame based on the previous frames, then uses that output as input for the next prediction. This continuous loop allows the model to generate video in real time while maintaining coherence and quality.

The technical challenge was converting Runway's non-causal architecture into a causal one. The company's research and engineering teams had to fundamentally retrain the model so it could operate with only backward-looking context. The payoff is that GWM-1 can generate continuously for more than 40 minutes without meaningful quality degradation, with character faces remaining stable and no morphing issues across extended interactions.

What Can Characters Actually Do?

Characters are customizable AI avatars that can be created from a single image of a person or object. Users can add documents to ground the character's responses in specific knowledge, and set up tool-calling frameworks so the avatar can interact with external systems. The system works across a variety of animated styles, including realistic human avatars, illustrated characters, objects, and masks.

When you interact with a Character, the technical flow happens in real time. Your voice is processed by an audio agent, passed to the model, which generates both video and audio via diffusion in latent space (a compressed mathematical representation). A second model then decodes that output into pixels and sound, which are sent back and rendered on your screen. The entire process, from speaking to seeing the response, takes less than 1.75 seconds end-to-end.

How to Deploy AI Characters in Production Environments

• Single-Image Customization: Create a working Character from just one photograph of a person or object, eliminating the need for extensive character design or animation assets upfront.
• Knowledge Grounding: Add documents and context to the Character so it responds based on specific information, making it suitable for customer support, product training, or domain-specific applications.
• Tool Integration: Set up tool-calling frameworks that allow Characters to interact with external systems, databases, or APIs, enabling them to perform actions beyond conversation.
• Style Flexibility: Deploy the same underlying model across different visual styles, from photorealistic avatars to illustrated characters, without retraining.

Why This Matters for AI Product Development

Characters represents a shift in how research, engineering, and product teams collaborate at AI companies. Rather than product teams deciding what to build and asking research to figure out how, Characters emerged because of a research breakthrough in GWM-1. The model's capabilities enabled the product, not the other way around.

This inversion of the traditional product development process reflects a broader challenge in generative AI: research breakthroughs often arrive before clear product applications exist. Companies like Runway are learning to build products around what their models can actually do, rather than forcing models to fit predetermined product requirements.

The most significant technical achievement is that previous AI avatars failed not because they looked wrong, but because they moved wrong. Simulating human expressions and movements has proven surprisingly difficult because humans are uniquely attuned to subtle facial and body language cues developed over millions of years of evolution. Runway's GWM-1 appears to have solved this problem, with Characters displaying natural movement and expression that doesn't trigger the uncanny valley effect.

What Does This Mean for the Broader Video AI Landscape?

Runway's shift to real-time generation comes as other companies pursue different approaches to video generation. ByteDance's Seedance 2.5, announced on June 23, 2026, takes a different path by generating 30-second clips in a single offline pass without stitching, addressing the problem of visible seams when clips are joined together. Both approaches solve different production problems: Runway enables live interaction, while Seedance enables longer, seamless pre-recorded sequences.

The emergence of multiple architectural approaches suggests the AI video market is maturing beyond a single dominant paradigm. Real-time models like GWM-1 will likely dominate interactive applications like customer support, live streaming, and gaming, while offline models like Seedance 2.5 will remain the standard for commercial advertising, film production, and content creation where pre-recorded quality matters more than latency.

For businesses considering AI avatar deployment, the key question is whether they need real-time interaction or pre-recorded quality. Runway Characters enables use cases that were impossible before, from live customer support avatars to interactive educational experiences. The technical achievement of moving from offline to real-time generation represents a genuine capability expansion, not just an incremental improvement in existing workflows.