Runway AI video bug became a product feature for its avatar model

Runway turned a Runway AI video bug into a user-facing feature after weeks of trying to remove drift in its real-time avatar model, according to Ryan Phillips, the company’s head of enterprise product. The episode, described at VB Transform 2026, shows how applied AI teams are still shipping around model limits rather than waiting for clean fixes inside the model stack.

Runway builds generative video tools based on what it calls general world models. Phillips used Runway Characters, a real-time video system for interactive AI-generated avatars, as the main example of how the company evaluates and releases products with nondeterministic outputs.

What was the Runway AI video bug?

Phillips said Runway’s real-time avatar system had a persistent problem after model optimization: characters could sway or move away from the center of the frame during generation. The team found that videos stayed more stable when the starting image was centered precisely.

Rather than keep spending time on a back-end model change, Runway added a front-end option called “Optimize for Image Quality.” The feature automatically recenters a user’s input image before generation begins, reducing the visible drift. Phillips framed the move as a way to convert an internal model constraint into something customers experience as a helpful workflow control.

How Runway says it evaluates real-time AI video

Phillips told the audience that evaluation starts with a shared definition of quality across product, design, research and sales, not only engineering. He said Runway runs internal reviews of generated examples so teams can agree on what counts as acceptable output and what counts as failure.

The evaluation set includes common customer scenarios and edge cases. Phillips cited a non-human character called “Tooth,” which has no nose and unusual teeth, as a test case for behavior outside typical human face structure.

Runway grades subtle artifacts strictly, according to Phillips. A generation can fail even if a face remains coherent, if background elements such as a net begin to warp. The tracking system is not exotic: Phillips said the company logs daily tests in an Excel spreadsheet, classifies failures as minor or major and ships when the model clears a pass threshold set in advance.

Phillips also said language models can help automate visual review, especially for spotting morphing or other changes in evaluation outputs. He described LLM-based judging as a way for teams to reduce manual review work, while still giving models additional context such as sketches or ad layouts.

Why the model optimization created product risk

Runway’s latency work begins with a large foundation model, then uses distillation to train a smaller model to imitate it. Phillips said that step cuts generation time by 80% to 90%. The company then applies adversarial post-training, which tests the model against systems designed to expose weaknesses and helps recover visual quality lost during distillation.

Those steps also introduced the avatar drift issue, according to Phillips. For teams building real-time AI products, the trade-off is familiar: faster inference can change output behavior, and product teams may have to decide whether to keep tuning the model or alter the interface around the defect.

Infrastructure problems also shaped the release

Phillips said Runway Characters also exposed lower-level infrastructure issues. After launch, 8% of API calls dropped to 16 frames per second, below the 24 frames per second target, causing visible stutter for customers.

Runway traced the issue using a Claude-powered AI agent with Datadog and Sentry, according to Phillips. The problem was isolated to one data center in the us-east-1 region. He said the fix was physical GPU replacement rather than a software configuration change.

Phillips described the development process as a period of repeated failure before improvement, using the phrase “failure hell.” He also argued that real-time generation will change creative work from making individual assets toward defining parameters, aesthetics and intellectual property that agents or video models can use to generate many outputs.

This story draws on original reporting from VentureBeat.