A Generated Frame Fails at the Wordmark First

Sit through a review where generated frames are being signed off and you learn quickly what actually gets them rejected. It is rarely that they look fake. It is that the brand mark is wrong. The letterforms are nearly right, a counter is the wrong shape, one terminal curls where the real one is cut flat.

Nearly right is worse than absent, because an audience does not read it as a mistake. It reads as a counterfeit of the brand's own mark. This is not a hypothetical failure mode — frames get rejected in our own review for exactly this, and the tell is almost always a letterform that passes a glance and not a second look.

This is not a prompting problem

An image model produces pixels that are statistically consistent with what it has seen. Typography is not a texture. It is a closed set of exact shapes carrying a specific identity, and a particular wordmark is one shape among millions that the model has no reason to hold precisely. So it renders the impression of typography: the right weight, the right rhythm, the wrong letters.

Which means the failure does not respond to better prompting. A sharper prompt produces a different wrong mark, faster.

The fix is a pipeline rule

The rule that solves it is blunt: the mark is never generated. It arrives as vector from the brand library and is composited on top of a generated plate — Photoshop for a still, After Effects or Nuke for motion. The same applies to every part of a frame that a proofreader or a lawyer would check.

• The wordmark and every lockup, in vector, from the source file.
• Packaging copy and regulatory panels, set properly rather than implied.
• The product's own geometry, wherever the product is recognisable.
• Any real person's likeness, which is a rights question before it is a craft one.

Provenance, so a question a year from now has an answer

Generated material is not something to hide. It is a fact to record. Per frame, keep what was generated, which tool produced it, and what was captured or rendered instead. Keep that record attached to the asset rather than in somebody's memory, and carry Content Credentials where the platform supports them. The reason is practical: a campaign outlives the team that made it, and the question of how a frame was made always arrives after everyone has moved on.

Disclosure is a decision to make early

Platforms are converging on labelling requirements for synthetic media and the specifics keep moving. The stable position is to decide the brand's own disclosure line before a campaign runs, rather than in response to being asked. A brand that says plainly how a frame was made gives up very little. A brand that is asked and cannot answer has already lost the argument.

Perspective slot — awaiting Belal's own words on how you spot a wordmark that is nearly right, and why nearly right is the dangerous version.

Put the mark check where the colour check already is

Most studios already have a habit for colour: someone opens the file beside the reference on a calibrated display before anything ships. Brand marks deserve the identical habit. Someone opens the generated frame next to the brand file at full size, looks at the counters and the terminals, and puts their name against it. That step belongs in the workflow, named, and not in the recovery afterwards.

It takes a minute. Taking a live campaign down takes considerably longer.

AI does not lower the standard for brand assets. It moves where the standard has to be enforced — off the set and into the review — and a studio that has not moved its review to match is running on luck.