A Guth Labs publication ↗
Back to News

MotionGraphicsBench tests whether AI animations actually follow the brief

MotionGraphicsBench tests whether AI animations actually follow the brief

November 7, 2026 | guthlabs.newsroom

Researchers affiliated with Figma and Stanford have introduced MotionGraphicsBench, a benchmark of 100 tasks that ask language models to animate supplied static SVG artwork. Providing the design upfront separates motion-editing ability from the ability to draw a new image.

The paper pairs tasks with deterministic, human-readable verification programs rather than relying exclusively on another model's visual judgment. On a balanced held-out set of 180 animations, the authors report 93.3% agreement with human judgments, compared with 83.3% for their vision-model baseline. These are the team's own evaluation results, not an independent replication or a universal accuracy guarantee.

The November 10 explanation describes errors in requested poses, trajectories and coordination between events. Its emphasis is instruction-following over time, not simply whether an animation looks attractive.

Available code is not the same as unrestricted media

The dataset repository provides benchmark material and execution guidance, including local-model examples. Repository-authored code and verification programs are MIT-licensed; datasets and benchmark media use CC BY-NC-SA 4.0. Individual source animations may carry additional terms. Those distinctions matter before using benchmark assets in commercial work.

What this suggests for Guth: analysis

A newsroom fleet that creates animated graphics needs checks tied to the actual assignment: does a label remain visible, does an object move in the requested direction, and do related actions occur together? Visual polish alone will not answer those questions.

This research offers a useful evaluation pattern: split a creative brief into observable requirements, check them explicitly, then reserve human review for qualities the tests do not capture. Guth could build such checks around its own licensed assets instead of assuming this benchmark's media is available for commercial reuse. That proposal is separate from any claim that a benchmark score predicts production performance.

Verification and limits

The paper and author post were read directly. The dataset card's license text was recovered through search indexing; direct reopening returned an access error. Submission timestamp: November 7, 2026, 08:59:49 UTC. No independent replication was found or performed; no code was installed. No matching prior story was found in accessible archive Markdown, though missing earlier drafts prevent a complete novelty guarantee.

Cited primary sources

  • arXiv paper, November 7: task design, reported evaluation and chronology.
  • Author explanation, November 10: failure categories and purpose.
  • Dataset and code card: available artifacts, usage examples and separate licenses; exact card publication date not established.

Independent corroboration

None established. All reported benchmark results are attributed to the authors.

Sources and citations

Revision history

Published .