◉ Learn

Deterministic brand memory: constraints, not prompts.

Deterministic brand memory is a machine-readable definition of a brand, applied as a constraint at composition time rather than as a suggestion in a prompt. Logo, palette, type and safe zones come out identical on every asset. ElarisLabs uses it to hold a brand fixed across hundreds of variants and every language.

This page covers why generative tools drift off-brand, what has to be locked, what should stay generative, and how brand memory differs from a brand kit, a prompt library and a fine-tune.

Last updated 11 September 2026

◉ The problem

Generative tools re-interpret a brand on every run.

Generation is probabilistic by design. That is the feature: ask for a kitchen scene ten times and you want ten kitchens. The same property applied to a logo gives you ten logos, and at ad sizes the difference between them is not subtle.

It rarely shows up on the first asset. It shows up at volume, when a brand manager lays 200 outputs on a wall and finds four blues, three logo proportions and a headline hierarchy that wandered. At that point the review cost has moved from one master to every single file, which is the cost the tool was bought to remove.

BRAND IN THE PROMPTPrompt"use our blue, our logo"Modelrun 1run 2run 3three different blues,three redrawn logosBRAND AT COMPOSITIONPromptscene onlyModelBrand memorylogo placed, palettepinned, type lockedrun 1run 2run 3identical brand elements, every run
The only structural difference is where the brand enters. Ask the model for it and you get a likeness with variance. Apply it after generation and the variance has nowhere to land.

◉ The line

What gets locked, what stays generative.

Locking everything defeats the purpose of a generative system. The line sits between what the brand team approves once and what the campaign is meant to vary.

Locked, by value

  • Logo lockup and clear space

    A regenerated logo is a redrawn logo. It is placed, never generated.

  • Palette, by exact value

    Near-enough brand colour is off-brand colour, and it compounds across a set.

  • Type hierarchy and the Arabic face

    Weight, size ratio and the right script face, so every language matches the approved master.

  • Grid, margins and safe zones

    The composition rules that survive a change of frame.

  • Legal and mandatory copy

    Disclaimers, terms marks and market-specific lines that must appear verbatim.

Left generative

  • Backgrounds and scene

    Where model quality actually earns its keep.

  • Product staging

    Per SKU, against the same approved composition.

  • Headline and body variants

    Within the tone rules, for testing.

  • Motion and pacing

    Within the brand's timing and transition rules.

◉ The boundary

Three things it is not.

Most tools that claim brand support are doing one of these. Each moves the odds. None of them makes the output identical.

Not a PDF brand guideline

A guideline is written for humans to interpret. A generative system cannot read intent out of a PDF, so the same document produces a different result every run. Brand memory is the same information in a form the composition step can enforce.

Not a prompt library

Prompts are suggestions with a probability attached. Repeat the identical prompt and the output moves. Anything that must be identical across 500 assets cannot live in the prompt.

Not a fine-tune

Training on brand assets shifts the average output towards the brand. It does not guarantee the logo is the logo. Fine-tuning changes the odds; brand memory removes the question from the generative step entirely.

The test for any tool: ask where the brand kit is applied. Before generation is a preference. After generation, at composition, is governance. Only the second one holds when the same master has to ship in Arabic and English across a full placement matrix.

◉ FAQ

Brand memory, answered.

What is deterministic brand memory?

Deterministic brand memory is a machine-readable definition of a brand, covering logo lockups, exact palette values, type hierarchy, grid, safe zones and mandatory copy, that an AI creative system applies as a constraint at composition time rather than as a suggestion inside a prompt. Because the constraint is applied after generation, the same brand elements come out identical on every asset. ElarisLabs uses it to hold a brand fixed across hundreds of variants and across languages.

How do I stop AI creative drifting off-brand?

Move the brand out of the prompt and into the composition step. Generation is probabilistic, so anything you ask a model to reproduce, such as a logo, a specific hex value or a type hierarchy, will vary run to run. The fix is to let the model generate only what is allowed to vary, such as backgrounds, staging and scene, and to place the fixed brand elements deterministically on top under rules the system cannot violate.

Why do generative tools produce a slightly different logo every time?

Because they are drawing it, not placing it. A diffusion model reconstructs a logo from learned statistics, so proportions, spacing and letterforms shift on each run, and at small sizes the drift is obvious. Any system that treats the logo as a placed asset with defined clear space rather than as something to generate avoids the problem entirely.

What should stay generative?

Everything whose variation is the point: backgrounds, scene composition, product staging per SKU, copy variants inside the tone rules, and motion within the brand's timing rules. Locking those removes the reason to use a generative system at all. The line sits between what the brand team approves once and what the campaign is meant to test.

Does brand memory hold across languages?

It has to, or the localised campaign becomes a second approval cycle. The brand kit carries a script-appropriate typeface for each language it supports, so an Arabic version inherits the approved hierarchy rather than falling back to a default font. The mirrored layout still has to follow the brand grid and clear space rules, which is the part covered in RTL ad localisation.

How is this different from an AI tool that says it supports brand kits?

Ask where the brand kit is applied. If it is fed into the prompt or used to bias generation, the output is a likeness that varies. If it is applied as a constraint at composition time, with exact values and placement rules, the output is identical every run. The first is a style preference, the second is governance, and only the second survives a legal review at volume.

Put your brand under lock.

Load the kit once. Every asset after that inherits it, in every language.

100%
brand IP compliance
Weeks→Min
production velocity
formats per creative
1
source of truth

◉ Scale without the headcount

One canvas.

Infinite assets.

No credit card · unlimited seats on trial · go live in minutes