Marketing needs twelve social-launch variants by Thursday, the brand guidelines were last touched in 2023, and you have three image models open in three browser tabs. You could generate forty assets in an hour. The problem is not volume; the problem is that an image model is an averaging machine with a strong accent, and if you let it run unsupervised across a launch, your brand quietly dissolves into the model's house style. Firefly centers everything and lights it like a perfume ad. Recraft flattens and geometricizes. Ideogram nails typography and then drifts the palette. This lesson teaches you to generate brand assets across three models without flattening your visual identity into theirs, to run each output through a brand-consistency review that has named pass criteria, and to walk away with a publishable social pack of the three variants that held plus a model-fingerprint note you commit to the brand library so the next designer does not relearn this the hard way.

The Launch Pack That Looked Like Three Different Brands

Start with the failure you are trying to avoid. A brand designer is handed a product-launch ticket: twelve social variants, four formats, one campaign. They are smart, they are fast, and they distribute the work across the three image models they trust most. Four go to Adobe Firefly because it carries commercial indemnification and legal sleeps better. Four go to Recraft because it does clean vector-flavored graphics. Four go to Ideogram because the campaign has typography baked into the image and Ideogram is the one model that can render readable words.

An hour later they have twelve assets. Each one, in isolation, is good. Laid out together in the campaign deck, they look like three different companies sharing a hashtag. The Firefly assets are warm, centered, softly lit, with that faint cinematic vignette the model loves. The Recraft assets are flat, high-contrast, geometric, almost Swiss. The Ideogram assets have crisp headline type and a color palette that has drifted two steps toward the model's defaults. The brand had one voice on Monday. By Thursday it has three accents, and the campaign reads as incoherent to exactly the audience a launch is supposed to convince.

The instinct is to blame the designer for not prompting harder. That is the wrong lesson. The right lesson is that each model has a fingerprint, the fingerprint is strongest exactly where the brief is thinnest, and the only defense is a review pass that judges every asset against the brand and not against itself. You do not fix this with a better prompt. You fix it with a process that assumes drift and catches it.

Why Every Model Has an Accent You Cannot Prompt Away

An image model is trained on an enormous corpus, and the statistical center of that corpus becomes the model's default aesthetic. When your prompt is specific, you pull the output toward your intent. When your prompt is vague, or when you ask for something the model has weak data on, the output falls back toward that center. This is why the accent shows up most on the easy assets, the background-and-logo social tiles where you did not bother to over-specify, and least on the hero shot you art-directed for twenty minutes.

Firefly's center is commercial-photography lighting and centered composition, a legacy of its Adobe Stock training base. Recraft's center is flat, illustrative, vector-adjacent graphics with clean edges, which is why it is the model people reach for when they want something that could become an SVG. Ideogram's center is typographic competence, the one thing most image models fail at, paired with a tendency to pick its own saturated palette unless you pin the colors. None of these are flaws. They are the reason you chose each model. They become flaws only when they overwrite your brand instead of serving it.

The Three Models and What Each Is Actually For

Treat the three not as interchangeable generators but as three specialists with three jobs. Firefly is your safe-for-client photographic and lifestyle imagery, the assets that need a person, a product in a real setting, or soft editorial warmth, and the assets a nervous legal team has to sign off on, because the indemnification on paid Creative Cloud plans is real and it matters when the asset is going out under the company's name. Recraft is your graphic, illustrative, and near-vector work, the bold flat compositions and icon-flavored tiles. Ideogram is your typographic hero, the asset where the headline lives inside the image and has to render as crisp readable type rather than the melted pseudo-letters most models produce.

Picking the right specialist per asset is half the battle. The other half is making three specialists sound like one brand, and that is what the review pass is for.

The Twelve-Variant Generation Pass, Done Deliberately

Here is the workflow that produces twelve usable variants instead of twelve accents. Before you open a single model, write the brand constraints down as an explicit block you paste into every prompt: the exact hex values of the primary and secondary palette, the typeface or its closest available substitute, the compositional rule the brand follows (left-aligned, generous negative space, off-center subject, whatever it actually is), and the tonal words that describe the brand voice in images (calm, precise, human, not playful, not corporate). This block is the same in all twelve prompts. The only thing that changes per asset is the subject and the format.

Then assign each of the twelve to its specialist. The four lifestyle and product tiles go to Firefly with the constraint block plus the subject. The four graphic and pattern tiles go to Recraft with the same block. The four headline tiles go to Ideogram with the block plus the exact headline string in quotes so it renders the words you wrote, not a paraphrase. Generate two or three options per asset, not one, because you are about to reject most of them on purpose.

The discipline here is that you are not trying to get twelve keepers. You are trying to surface which assets each model can hold the brand on and which it cannot. Generation is cheap. Selection is the craft. Expect to generate thirty-plus options to keep three, and treat that ratio as healthy, not wasteful.

The Brand-Consistency Review, With Named Pass Criteria

This is the heart of the lesson. A brand-consistency review is not "do I like it." It is a fixed set of pass criteria you run against every generated asset, the same way an accessibility audit runs a fixed set of checks. Vague review produces vague results; you nod at things that drift because nodding is faster than rejecting. Named criteria force the rejection.

Run each asset against these five checks, and an asset passes only if it clears all five.

The Five Brand-Consistency Checks

1. Palette lock. Eyedrop the dominant colors. Do they sit on the brand palette, or has the model drifted toward its own defaults? A two-step hue shift toward the model's preferred saturation is the most common and most invisible failure. If the brand blue came back as a slightly more purple, slightly more saturated blue, it fails.
2. Compositional rule. Does the asset follow the brand's actual layout logic, or did the model impose its center-everything reflex? If your brand is built on asymmetry and generous negative space and the model returned a centered, balanced, filled composition, it fails even if it is beautiful.
3. Tonal match. Cover the logo and ask whether a stranger would attribute this to your brand or to "generic premium SaaS." If the lighting, mood, and styling read as the model's house style rather than your brand's voice, it fails.
4. Typographic integrity. If the asset contains words, are they your exact copy in your typeface, rendered as real letters, with correct spelling and no melted glyphs? Models hallucinate text constantly; an asset with a single garbled character fails outright because it cannot ship.
5. Format and safe-area fitness. Does the composition survive being cropped to the platform's aspect ratios and overlaid with the platform's UI chrome? A hero subject dead-center of a 1:1 tile that gets cropped to a 9:16 story and loses its subject fails.

Run the twelve-plus assets through these five checks and a brutal thing happens: most of them fail at least one. That is the system working. You are not looking for assets you like. You are looking for assets that hold the brand under the model's pressure to flatten it.

An image model does not flatten your brand because it is bad. It flattens your brand because flattening toward its training-set average is the only thing it knows how to do. The review is where you put the brand back.

A Worked Example: One Launch Tile, Three Models

Make this concrete. The campaign is a product launch, and one of the twelve assets is a square social tile reading "Now shipping" over a hero treatment of the product on the brand's signature warm-neutral background. You decide, before opening anything, that this tile has text baked into the image, which makes it an Ideogram job by your specialist routing. But you also want to see what the other two do, partly to test and partly to populate the fingerprint note, so you run it through all three with the identical constraint block.

Ideogram returns the cleanest type, the headline crisp and correctly spelled, but the warm-neutral background has drifted toward a more saturated peach because Ideogram pinned the palette poorly. It passes typographic integrity, fails palette lock until you correct the background hex. Firefly returns a gorgeous, soft, centered composition with the product glowing in studio light, but it rendered the headline as melted pseudo-letters, an automatic typographic-integrity fail, and it centered everything against the brand's off-center rule, failing the compositional check too. Recraft returns a flat, high-contrast, almost poster-like tile that is striking but reads as a different, bolder brand than yours, failing tonal match. The exercise tells you exactly what the assignment already implied: Ideogram is the right specialist for this tile, and the other two each failed in their signature way, which is precisely the data the fingerprint note exists to capture.

So you keep the Ideogram version, correct the background to the exact brand hex, and move on. The whole comparison took ten minutes and it did two jobs: it confirmed the right tool for this asset, and it produced three fresh fingerprint observations (Ideogram peach-drift, Firefly melted-text-on-typographic-assets, Recraft over-contrast) that make the next launch faster. This is what the workflow feels like in practice: not slower than reckless generation, but more deliberate, and compounding.

Keeping the Three That Hold

After the review you will have a handful of survivors. The lesson's named artifact is a social pack of the three best, one from each model ideally, because a pack that proves all three specialists can be brought into one voice is worth more to the brand library than three Firefly assets that happened to pass. Three is the target not because three is magic but because forcing yourself to three forces the rejection discipline; a designer who keeps eight is usually keeping the ones they are tired of fighting, not the ones that earned it.

For each of the three survivors, do the last ten percent of craft the model cannot. Nudge the type to the real baseline grid. Replace the model's near-brand color with the exact brand hex using a fill layer if the eyedropper caught a two-step drift you decided to ship-and-fix. Composite the real logo at the real clear-space margin rather than trusting the model's rendering of it. This hand-finishing is what separates a generated asset from a brand asset, and it is fast precisely because you are doing it on three assets instead of twelve.

The Model-Fingerprint Note for the Brand Library

The second half of the artifact, and the part that compounds, is a model-fingerprint note. This is a short document you commit to the brand library that records, for each of the three models, what its accent is, where it drifts, and the prompt constraints that pull it back toward the brand. It is the institutional memory that means the next launch costs ninety minutes instead of six hours.

A fingerprint note entry reads like this. For Firefly: tends to center and over-light; pin composition with explicit "off-center subject, left third, generous negative space"; carries commercial indemnification on paid plans so prefer it for any asset shipping under the company name; drifts warm so pin the brand-neutral white balance. For Recraft: flattens to vector-clean by default which suits our graphic tiles; drifts toward higher contrast than the brand uses, so specify the exact muted contrast; strong choice when the asset may later become an SVG. For Ideogram: the only reliable typographic renderer of the three; always supply the exact headline string in quotes; pins the palette poorly, so lead the prompt with the hex values and re-check palette lock every time.

The note is not a one-time document. It grows. Every launch you add the drifts you found and the constraints that fixed them, and over a few quarters it becomes the single most useful page in the brand library, because it encodes the one thing no model's documentation will ever tell you: how this specific model behaves against your specific brand.

The Legal Layer: Indemnification as a Routing Input

There is one more reason the routing matters, and it is not aesthetic. When an asset ships publicly under the company's name, it carries legal exposure, and the three models do not carry the same protection. Adobe Firefly offers commercial indemnification on paid Creative Cloud plans, which is why a cautious legal team will sign off on a Firefly-generated hero faster than on output from a model with no such cover. This is not a small detail in 2026, with the Disney, Universal, and Warner Bros. litigation against Midjourney still live and every brand legal team newly alert to where generated imagery came from.

Treat indemnification as a routing input alongside aesthetics. For the assets that ship publicly and carry the most exposure, bias toward the indemnified model even when another model's output is marginally prettier, because the marginal aesthetic gain does not justify uncovered legal risk on a public launch asset. For internal, exploratory, or heavily-recomposed work where your own composition is the final artifact, the non-indemnified specialists are appropriate. This pairing of aesthetic fit with legal exposure is exactly the kind of judgment that separates a brand designer from a button-presser, and it belongs in the fingerprint note too: which model you trust for public assets and why.

Why This Protects Both the Brand and the Designer

It is tempting to read this as a lesson about quality control, and it is, but the deeper point is about leverage. A brand designer who can run three models against one voice, reject most of the output on named criteria, and document the fingerprints is not slower than the designer who generates twelve assets in an hour. They are faster across the quarter, because the fingerprint note collapses the next launch, and they are safer, because the review catches the drift that would otherwise reach the audience under the company's name.

The designer who gets hurt in 2026 is the one who treats image models as interchangeable and trusts the output because it looks finished. The designer who thrives is the one who treats each model as a specialist with a known accent, builds a review that assumes flattening, and turns every launch into a slightly better-documented brand library. The volume the models give you is real. The voice is yours to protect, and this is how you protect it without giving up the speed.

Putting It to Work This Week

On your next launch ticket, do four things. Write the brand constraint block once, before you open any model, and paste it into every prompt unchanged. Assign each asset to its specialist model on purpose, not by habit. Run every output through the five named brand-consistency checks and keep only the assets that clear all five, aiming for three survivors. And start the model-fingerprint note, even if the first version is three bullet points, because the version that exists is worth more than the perfect version that does not. The pack you ship is the visible deliverable. The fingerprint note is the one that makes you faster every launch after this one.

Key Takeaways

• Every image model is an averaging machine with a strong accent: Firefly centers and over-lights, Recraft flattens to vector-clean, Ideogram renders type well but drifts the palette. The accent overwrites your brand most where the brief is thinnest.
• Treat the three models as specialists with assigned jobs - Firefly for indemnified lifestyle and product imagery, Recraft for graphic and near-vector tiles, Ideogram for typographic headlines - rather than interchangeable generators.
• Write the brand constraints (exact hex palette, typeface, compositional rule, tonal words) as one block and paste it unchanged into every prompt; only the subject and format change per asset.
• Run every generated asset through five named brand-consistency checks - palette lock, compositional rule, tonal match, typographic integrity, format and safe-area fitness - and keep only assets that clear all five. Expect to generate thirty-plus to keep three.
• Hand-finish the three survivors (baseline grid, exact brand hex, real logo at real clear space); this last ten percent is what turns a generated asset into a brand asset and is fast because you do it on three, not twelve.
• Commit a model-fingerprint note to the brand library recording each model's accent, drift, and corrective constraints. It is institutional memory that turns the next six-hour launch into a ninety-minute one and compounds every quarter.