
Every week there is a new demo reel that makes AI video look solved. It is not. Anyone who works with these tools for a living knows where they fail, and a studio that cannot tell you is either inexperienced or selling you something. Here is our field report, set against what recent benchmarks are finding.
Text inside the picture
Models still struggle to spell. A benchmark published on 1 October 2026, VTR-Bench, found that the best video model it tested had a word error rate of 0.25 when rendering on-screen text, which is roughly one word in four wrong. On a pack, a shop sign or a title, one wrong word is a brand error.
What we do: we add titles, pack details and logos in the edit rather than asking a model to generate them, and we keep brand colours and logos exact.

Precise, multi-step action
Pouring, unwrapping, stirring, opening: anything where a hand must do several things in order is hard. A benchmark of 110 goal-directed manipulation tasks, Ego2Act, found dropped steps and physics breaks even in strong models.
What we do: we plan these shots as sequences of simpler ones, generate several attempts and keep the best take, the way an editor selects from a day’s rushes. Where the action must be exact, live footage inside a hybrid film is often the better answer.
Drift over time
Faces, costumes, rooms and light shift between shots, and over a long clip they can shift within one. Even as clips get longer, stability over long durations remains a known weak spot, a point that coverage of the open LTX-2 model made when it launched.
What we do: drift is the problem our whole pipeline is designed around. Character and environment sheets are approved before video exists, every shot starts from an approved keyframe, and a creative director reviews every shot against the sheets.

The small tells
Extra fingers, objects that float, hair that merges with a collar, crowds whose faces dissolve at the edges. These are less common than they were, and they have not gone away. QA is now a recognised stage of AI production, not an afterthought.
What we do: our review asks the questions a continuity supervisor asks. Does the face match in the wide and the close-up? Is this the right costume? Does the light match the shot before? Are buildings and props in proportion?
What this means when you buy
Ask to see work that has run, not tests. Ask what was fixed in the edit and what was regenerated. Ask how a bad shot is caught, and who catches it. A studio that talks easily about its failure modes has usually met them.
Honest limits are part of expertise. If you want to talk through where they would bite on your brief, write to namaste@teevra.co.


