How Runway turned a real-time AI video model bug into a product feature and what AI builders can learn

Runway's real-time AI video model for avatars had a stubborn center drift bug. Instead of patching the core model, the team shipped a frontend feature that re-centers the input image before generation. The story, shared by Runway ML head of enterprise product Ryan Phillips at VB Transform 2026, offers practical lessons for any team building real-time AI experiences: turn model limitations into product features, define quality cross-functionally, and invest in full-stack observability.

What happened

Runway Characters is a real-time video model that enables zero-latency, interactive AI-generated avatars. To achieve real-time latency, Runway used distillation to reduce generation time by 80-90%, training a smaller student model to mimic a larger teacher model. They then applied adversarial post-training (APT) to regain visual sharpness lost during distillation. However, these architectural changes introduced a new problem: characters would sway or drift from the center of the frame during real-time generation.

The team spent weeks trying to fix the core model. They eventually discovered that if the user's initial input image was perfectly centered, the generated video remained stable. Instead of continuing the backend engineering effort, Runway introduced a frontend feature called "Optimize for Image Quality" that automatically re-centers the user's image before generation begins. Users perceived a helpful feature rather than an engineering flaw.

Why AI builders should care

Runway's approach to evaluation is as instructive as the bug workaround. The team runs cross-functional workshops with product, design, research, and sales to define what "quality" looks like for real-time generation. They build an evaluation set covering broad customer use cases and extreme edge cases (like a non-human character called "Tooth") and track pass/fail rates in a simple Excel spreadsheet. They set a pass rate bar before shipping. For teams dealing with non-deterministic quality drift, Phillips noted that LLMs can automate visual grading of morphing or changing artifacts.

Practical implications

Three concrete takeaways for AI builders:

• Frontend workarounds can ship faster than backend fixes. If a model limitation is consistent and detectable at input time, wrapping it in a UX feature can reduce time-to-ship without changing the core architecture. The center drift bug became "Optimize for Image Quality" instead of a months-long research project.
• Full-stack observability is non-negotiable for real-time AI. After launching Runway Characters, 8% of API calls dropped to 16fps. The team used a Claude-powered AI agent with Datadog and Sentry to trace the issue to a single data center in us-east-1. The fix was physically replacing GPUs, not changing config.
• Cross-functional quality definitions prevent shipping the wrong thing. Runway's evaluation workshops align the entire organization on specific failure modes before any model ships. The tool is simple (Excel), but the process is rigorous.

Caveats

This account is based on a single VentureBeat article describing Runway's internal processes and product decisions. The details may represent a selective company narrative rather than exhaustive technical documentation. The distillation and APT techniques are described at a high level without independent benchmarks or reproducibility details. The "Optimize for Image Quality" feature's effectiveness and user reception are not independently verified. Hardware and infrastructure details (Nvidia collaboration, GPU replacement) are presented as anecdotal evidence from one presentation.