Text-to-video in production: notes from the queue

The demo reels stopped being the interesting part. Kling, WAN, MiniMax and their generation of video models can hold a shot now — coherent motion, believable physics, water that behaves like water, which anyone who has watched this space for two years knows was the boss level. The interesting part is what happens when you put these models behind a product: the failure modes move from the pixels to the pipeline.

Budget in usable seconds, not generations

Video pricing reads per generation or per second of output, and both numbers are fictions until you multiply them. Real creative workflows keep maybe one clip in three to five — prompt drift, a weird hand, motion that dies halfway through. Your effective price is cost per second you actually ship, and it's several times the sticker. Two consequences:

• Premium models can be the cheap option. A model that costs twice as much per second but doubles your keep rate breaks even before you count the human time spent reviewing rejects.
• Iteration belongs upstream. Every discarded video generation is the most expensive way to discover your prompt was wrong.

The image-first workflow is the budget hack

The single most effective cost and quality lever we know: don't iterate in video. Generate a still first — text-to-image is one to two orders of magnitude cheaper — iterate on composition and look there, get approval on the frame, then hand the approved image to an image-to-video model to animate.

This is how the video on this site's landing page was made: a text-to-image pass for the framing, human selection, then image-to-video on the chosen still. The rejected iterations cost image prices. Only the approved compositions ever cost video prices. As a bonus, the workflow leaves a natural lineage trail — final clip, source frame, original prompt — which is exactly what provenance rules increasingly want you to be able to show.

Iterate at image prices. Commit at video prices.

Your architecture is now a job system, like it or not

A video generation takes anywhere from tens of seconds to several minutes, with heavy variance by model, resolution, and — the part nobody budgets for — provider queue depth, which swings with everyone else's launch day, not yours. Design consequences:

• Never hold a request open for a render. Submit, get a job id, deliver completion by webhook or poll. Every timeout between your user and the GPU is a place a six-minute render goes to die.
• Jobs need a real state model. Queued, running, succeeded, failed-with-cause — with terminal states that never un-happen. Media jobs are long enough that users navigate away, retry buttons get mashed, and idempotency stops being optional.
• Fetch the output immediately. Provider URLs expire — on one major provider in about an hour. Persistence is part of the job lifecycle, not a cleanup task.
• Queue variance is a failover trigger. A provider answering healthily but queueing ten minutes deep is, for your user, down. Cross-provider routing turns that from an apology into a routing decision.

Field notes on the current class

We run these models in production for the media on this site, so, briefly and without a leaderboard: the Kling line has been the consistent workhorse for image-to-video — strong motion coherence and the least prompt drift on liquid and cloth. WAN-class models are the value play for text-to-video drafts. MiniMax delivers distinctive camera movement. All of them rev fast — plan for the id you integrate to be renamed or retired within roughly a year, and keep your integration capability-shaped.