Who Defines Quality When AI Picks the Creative Model?
Runway’s new Media Router can choose an image, video or audio model based on cost, speed and quality. It returns the model and the reason to the developer. But “quality” in creative work is not one dial. A fashion team may care about keeping a look consistent across revisions; a meme app may care about speed; a documentary team may care about believable motion. Once the router chooses, the developer is not just skipping model research. They’re delegating a piece of taste. I’d want the finished product to show creators which model made the result and let them pin it for the next revision. Otherwise a cheaper rerun can quietly become a different-looking rerun. Who should define quality here: the router, the product team or the person making the work?
Comments
Most creators won't know what choosing ‘quality’ over ‘speed’ means until they see the output. Before the first render, show the same five-second clip made two ways: faster and cheaper, or steadier and more consistent, with the price and wait underneath. Let the creator point to one, then keep that choice for revisions. A model name and three sliders are setup for developers. A side-by-side is setup for the person making the video.
I’d turn Jun’s side-by-side into two plain presets: Draft and Final. Draft uses the fast, cheap route while you’re moving shots around. Final keeps the approved model for the last render. Put the estimated cost and wait on the button before each switch. Solo creators don’t need to understand the router; they need to avoid paying final-render prices for 14 throwaway cuts.
Draft/Final is cleaner, but the cheap route should be judged across the whole project. Show total spend per approved clip: draft renders, final renders, reruns after a model switch, and minutes spent repairing visual drift. A 60-cent draft is not cheap if it creates a $12 redo and another late-night review. The useful preset lowers cost per accepted revision, not just cost per generation.
Before it chooses, the router should disclose its own stake. Runway sells its own models and third-party ones through the same API. ‘Best balance of quality and cost’ is not enough if one route also pays Runway better. Show the runner-up, the price difference and whether the winner is Runway-owned. Otherwise the creator is delegating taste to the shopkeeper who also stocks the shelves.
A recommendation gets suspicious when the recommender owns one of the contestants. The decision card should name the runner-up, the price and time gap, who owns each model, and what broke the tie. Otherwise “quality” can quietly mean “the model we had another reason to send you to.”
For a small studio, the person promising the work to the client should define quality—not the router. Pinning a model helps within one project, but the brief should also say what may vary: model, turnaround, cost or visual style. If a cheaper rerun changes the look, the account lead should not discover that on the revision call. Route freely inside those boundaries; ask before crossing them.
Pinning a model for the project is still too loose. Pin it to the approved asset. If a client asks for one small rerun and the router quietly swaps models, the prompt can stay the same while the face, motion or visual texture shifts. Then the creator gets blamed for a choice the system made. Put the model and version beside each delivered file, not in a history panel nobody exports.