A brand kit survives a bad model week

A brand kit is the saved set of colours, fonts, logo, product images, tone, and default frame that generations are expected to obey without being reminded. When one model family has a bad week — queue, look, audio, a silent turbo you mistook for native dialogue — the kit does not go down with it. You change the catalog row. You do not rewrite the brand. Hex is still hex. The caption family is still the caption family. The logo file is still the logo file.

This is the kit-versus-model-week split. It is not the kit-versus-caption-preset mapping, and it is not a standing-rule versus hex store.

The kit is the part that is not allowed to be stochastic

Generation is a sample. A brand is the part that is not allowed to be. Anything that must be identical across runs has to be supplied as a constraint rather than hoped for. That is the entire argument for storing it once instead of pasting a paragraph into every prompt.

What belongs in the kit:

Fill, outline, accent hex.
Font, if you are actually shipping that typeface.
Logo file, not a description of a logo.
Product shots — references constrain a generated product harder than adjectives.
Default frame shape, so a vertical brand stops emitting accidental 16:9.
Tone, written as a brief for a copywriter: what it never says, not only what it sounds like.

What does not belong in the kit as a model pick:

"Always Veo."
A credit number.
A paragraph that says "cinematic, best model this week."
A face. Faces are reusable character assets or a consented twin, not a hex.

Overlay and caption styling is deterministic. A hex value is that hex value. Inside a generated image the same values are guidance. The glossary pitfall is treating a kit hex as a guarantee inside the render. It governs overlays and captions exactly. In a rendered scene it is a strong hint competing with the lighting the model chose. When the model has a bad week, that hint gets worse. The overlay does not. That is why the kit is the failover, not the prompt.

Caption Studio still has 45 CSS presets in nine groups of five. The kit maps onto one family. The family is the show. A bad Veo week is not a reason to jump from Classic to Hype because the plate came back neon. Contrast overrides are for readability. Taste overrides are how the series becomes a mood board.

A bad model week is a row problem

Catalog rows fail in boring ways. A family is slow. A look drifts. Native audio on one row is fine and the cheap turbo next to it is silent. None of those are brand problems until you store the row in the same place as the hex.

Versely rows, priced in credits:

Veo 3.1 — T2V, native audio, 4K tier, 64 credits, 4 / 6 / 8 seconds.
Seedance 2.0 — T2V and R2V, native audio, 1080p, 44 credits for T2V and 97 for R2V, 4–15 seconds.
Kling O3 Pro text to video — T2V, native audio, 56 credits, 3–15 seconds.
Kling 3 Turbo text to video — T2V, silent, 22 credits. Do not brief it as dialogue.

Google's Gemini Omni 1.1 Flash (27 August 2026) extends in 10-second increments to 40 seconds cumulative on Google's own surfaces. Versely's Omni model pages are Gemini Omni Flash Edit (50 credits, audio false) and Gemini Omni Video (29 credits, 4K, 4 / 6 / 8 / 10 seconds). Flux 3 Video (Black Forest Labs, 4 August 2026 GA) does up to 20 seconds with native audio on BFL's own terms; Versely's Flux 3 Text to Video is 34 credits, 5–20 seconds, native audio true. BFL's evals are BFL's, not Versely benchmarks.

When Veo is the wrong row this week, Seedance or Kling or Flux is a different row on the same bill. The kit still burns the same caption colour on whatever file comes back. That is the studio test: fail over without a new subscription and a new brand paragraph.

A vendor surface will not do this. Flow is how Gemini users run Veo. Kling's app is how you run Kling. Living in the surface means the brand kit is "whatever that vendor's default font is" on a bad week. Keep the model. Leave the surface.

What still holds when the generate does not

Captions. Burned-in captions are pixels. They do not care which model produced the plate. Transcribe, time, composite. Product names still need a human pass because transcription is where brands get misspelled. The kit font and the kit hex still bind. Caption styles is the catalogue. Pin the family for the series. Do not "try a more viral preset" because Seedance rendered a hotter background. That is a new show.

Safe area. Platform chrome still covers the bottom and the button rail. A bad model week does not move TikTok's UI. Keep the hook line out of the chrome. Recheck after every reframe.

Product photography. Store the shots in the kit. When the generate invents a seventh colourway, you still have the real pack to overlay, cut away to, or feed as a reference. Adjectives will not save a SKU. References might. A bad week on a text-to-video row is a reason to go image-to-video or reference-to-video off the kit stills, not a reason to rewrite the brand book.

Tone. "We never say clinically proven" is still true when the model is fluent. Check claims. Brand safety is a review layer. It is not an AI content label. Toggling YouTube's altered-content checkbox (YouTube Blog, 27 May 2026: photoreal overlay on Shorts; labels alone do not impact recommendations or monetization eligibility) does not verify the sentence on screen.

Content velocity is how much finished work ships per week. A kit is what keeps that throughput from becoming ten slightly different reds. Volume without a kit is faster drift.

Switch the row, keep the kit

The working pattern is dull, which is why it survives.

Kit first. Hex, type, logo, stills, default 9:16, tone. Saving the kit is not a generate.
Caption family second. One of nine, mapped to the kit, contrast-checked on a real plate.
Row third, per shot. Talking cinematic eight seconds is Veo. Fifteen-second native hold may be Seedance. Silent draft may be a turbo. Do not store that pick in the kit.
Fail over without a meeting. If the row is wrong this week, change the row. Recaption only if the timing broke. Do not recast the family.

The AI video generator is the launcher for the row. The AI caption generator is the burn. Neither object is the brand. The kit is.

If the week is a bad one on a family you usually like, keep the family in the archive and keep the kit in the account. Sign up and stop pasting hex codes into prompts that were supposed to be model switches.

FAQ

Should the brand kit name a default video model?

No. The kit is values that must not drift: colour, type, logo, stills, frame, tone. A model is a job pick. Defaulting the kit to one row is how a bad week on that row becomes a brand crisis.

If the generate ignores our hex, is the kit broken?

Not on the overlay. Kit hex binds on captions and graphics. Inside the rendered scene it is guidance. Fix the scene with references and a different row. Do not "fix" it by changing the brand colour every Thursday.

Do we recaption when we switch models?

If the new plate has different duration or different speech, yes — burn-in is pixels, and a trim means a recaption. If you only swapped a silent cut for another silent cut of the same length, the overlay may still fit. Preview. Do not assume.