Why AI Image Models Get Your Logo Wrong (and What Actually Fixes It)

Every marketer who has generated an image with their logo in it has seen the same thing. The scene is fine. The logo is almost right. A letter has an extra stroke, the wordmark is a shade too warm, the icon is 8 percent wider than it should be, the clear space is gone. Generate again and you get a different almost-right version. Ten runs later you own a family of near-misses and not one copy of your actual mark.

This post explains why that happens, what the model vendors themselves say about it, why “the newer model spells words now” does not solve it, what publishing a warped mark does to your trademark, and the one fix that works on every model. If you want the how-to for placement and sizing, that is in how to put your logo on AI-generated images. This is the why.

A logo is an asset, not a subject

An image model generates by reconstruction. It has learned what things look like from patterns in millions of images, and it produces a new image that satisfies those patterns and your prompt. That is why it is good at “a plumber in a van at golden hour”: there are a thousand valid plumbers and vans, and any of them will do.

A logo has exactly one valid answer. It is a file with fixed geometry, fixed letterforms, fixed color values, fixed spacing and clear space. There is no “close enough” because the entire point of a mark is that it is always identical. A generative model can only ever produce something logo-like, because reconstruction from patterns is the only operation it has. Graswald, a vendor that shipped a dedicated “logo repair” feature in 2026, put it plainly in their launch post: a generation model does not copy your logo, it reconstructs an image from patterns it has learned, and produces something logo-like rather than your logo. When a vendor ships a repair feature, the problem is not theoretical.

Why does fine detail suffer more than the rest of the scene? Diffusion models resolve an image from coarse to fine. Composition and large shapes are settled early; letterforms, thin strokes and small icons are settled last, in the steps where the model is least constrained and any error is already baked into the surrounding pixels. Text-in-image research describes the same mechanism from a different angle: models learn text as a visual pattern, not as a sequence of characters, so they replicate the look of a word without understanding it. A logo adds a layout problem on top: text on a curve, wrapped in a badge, paired with an icon, at a size where every pixel matters.

The three failure classes

It helps to name the failures separately, because the popular fixes address only one of them.

1. Letterform hallucination. Misspellings, doubled letters, extra strokes, a lowercase a swapped for a single-story a. This is the failure everyone notices and the one that has improved most. A 2024 example from a DALL-E 3 era write-up had “RETAINED EXECUTIVE SEARCH FIRM” come back as “RETAITED EXECUTIVE SEARCH FINM”, and every correction moved the letters somewhere new. Current models spell that phrase correctly most of the time. Most is the operative word.

2. Geometry and color drift. The letters are right but the proportions are not. The wordmark is stretched to fit the composition, the icon-to-text ratio shifts, the kerning tightens, the blue lands two hex values off because the scene’s palette pulled it warm. This failure is subtle at feed size and obvious in a brand audit, and it is the one reference images do least about, because a reference tells the model what to aim for, not how to measure.

3. Treat-as-style. The model absorbs the logo as a visual influence rather than an object. Logo-like shapes appear on the van, the mug, the shirt, the wall. The brand colors bleed into the scene. This failure gets worse when you push reference strength up, because you are telling the model the logo matters, and the model’s way of making something matter is to spread it around.

Reference images reduce class 1, do little for class 2, and can make class 3 worse. Prompt phrasing (“preserve the logo exactly”) nudges class 1 and 2 and cannot touch the underlying mechanism. Only one approach handles all three, and it is the last section.

What the vendors say in their own docs

None of this is a secret. The model makers document it, in their own words, once you know where to look.

• OpenAI. The GPT Image prompting guide in OpenAI’s developer cookbook tells users to spell tricky words such as brand names letter by letter, and when editing, to state invariants explicitly: preserve identity, geometry and layout. The guidance for logos in general is to create an original, non-infringing logo. You are being told the model treats brand names as fragile and expects you to guard geometry yourself. The ChatGPT Images 2.5 release was headlined on better preservation of the people and products in reference photos, which is a real improvement and also an admission of what came before.
• Google. The Gemini image generation docs describe reference images as up to ten images of objects with high fidelity, and one of their own examples asks the model to put a logo on a bottle. On Google’s own developer forum, a thread titled How to keep logo intact on a product image? reports that with Nano Banana the logos are changed regardless of the prompt insisting the product stay exactly the same. Google staff asked for screenshots. A separate Gemini community thread is titled, in full, “Nano banana pro fails to maintain the 100% consistency with the original photo”.
• Adobe. In a Firefly community thread on text and symbol generation, an Adobe community manager wrote that Firefly, and most other AI image generators, is generally not good at putting exact text from a prompt into an image, and recommended adding text afterwards in Express or Photoshop. Users in the thread reported a name misspelled one time in four and a web address rendered as a string of extra letters. Adobe’s generative AI guidelines also prohibit submitting inputs that include trademarks you lack rights to, and Firefly’s indemnity does not cover output generated from a prompt or reference that included a third party’s mark.
• Midjourney. The Omni Reference documentation notes that intricate details such as freckles or clothing logos may not match perfectly. Third-party guides recommend reference weights of 400 to 1000 for a corporate logo and, in the same breath, warn that weights above 400 can yield unpredictable artifacts. Midjourney has since replaced Omni Reference with an Edit Model in V8, which is worth remembering: the workaround keeps changing, the problem does not.

The pattern across all four is identical. Text is better. Preservation is a feature still being improved. The recommended fix for anything that must be exact is to add it afterwards.

Why the drift compounds

There is a second mechanism that makes this worse in real workflows: edits re-render. On OpenAI’s developer forum, users documented that asking for one change to a generated image re-drew everything, including objects, background and theme they had not mentioned, because the model cannot update an existing image based on its current contents; it produces a new image that resembles the old one. Every edit is a new draw, and every new draw is a new approximation of the logo.

So the practical arithmetic is bad. A campaign needs ten images. Each goes through two or three edit passes. That is twenty to thirty independent reconstructions of your mark, each with its own drift. One vendor writing about product-photo drift put it well: at fifty variants, that drift shows up dozens of times before anyone catches it. The brand-consistency guide for reference images covers how to keep products and people stable across a campaign. For the logo there is no stabilizing; there is only keeping it out of the model entirely.

The trademark angle marketers skip

A logo is not just a picture; it is the thing your trademark rights attach to. Two consequences follow.

Your own mark. Trademark protection rests on consistent use of the mark as registered. A stream of slightly different versions published under your own name blurs the distinctiveness you would need to enforce the mark against someone else. Nobody will sue you for warping your own logo. You are simply eroding an asset you paid to build, one feed post at a time.

Anyone else’s mark. If a partner logo, a customer logo, or a competitor’s product ends up in a generated image, distortion is not a defense. In Getty Images v Stability AI, the UK High Court in November 2025 rejected Getty’s copyright claims but found trademark infringement in instances where outputs carried distorted Getty watermarks. In the parallel US case, the Northern District of California in April 2026 declined to dismiss the Lanham Act claims, reasoning that distorted watermarks without attribution could mislead consumers about source or affiliation. The courts treated a mangled mark in AI output as a trademark question. That is the standard a marketing team should assume applies to a co-marketing image with a warped partner logo.

None of this requires a lawyer to act on. It requires the logo never to pass through the model.

The fix: generate the scene, place the mark

The only approach that handles all three failure classes is to separate the two jobs. Generation creates the scene. The approved source file creates the mark. A design agency that ranks for this exact question says the same thing in nearly the same words, and they are right.

1. Generate without the logo. Prompt for the scene, the subject, the mood, the composition. Never mention the logo. If the model does not know a logo is wanted, class 3 cannot happen.
2. Composite the real file after generation. A transparent PNG or SVG at full resolution, placed with rules rather than taste: same corner by default, sized to the image’s short side, margin of half the logo’s height, moved only when the corner is busy, padded only when contrast is poor. The placement guide has the numbers. Because the file is placed, not drawn, classes 1 and 2 cannot happen either.
3. Lock colors and fonts as assets, not words. “Use our blue” is a prompt the model will interpret. A hex value applied in the composite step is a fact. The same goes for any tagline or phone number under the logo: live text in your real font, added in the same step.
4. Keep the clean original for every edit. Any AI edit (upscale, reframe, inpaint, remix) runs on the version without the logo, and the logo is re-applied to the result. Edit the composited version and the model will treat the logo as part of the picture and redraw it, which puts you back at the start.
5. Put a human gate before anything ships. Inspect logos and product marks at 200 to 400 percent zoom, because class 2 failures are invisible at feed size and obvious on a billboard. The AI image QA checklist covers what else to check.

That is a harness, not a prompt. Marqeable is built this way: the model draws the scene, the real logo file is composited afterward with placement chosen by corner busyness and sizing by short side, brand colors and fonts are applied as locked assets from your brand look, the clean original is kept so every AI edit operates on it and the logo is re-applied to the result, and a person approves every image before it goes anywhere. The logo never passes through the model, so it is never wrong.

Frequently asked questions

Why does the AI keep changing my logo even when I upload it as a reference?

A reference steers the drawing; it does not paste the file. The model reconstructs the mark from learned patterns, so letterforms, proportions, spacing and color come out slightly different on every run. References reduce misspellings, do little for geometry and color drift, and can make the model treat the logo as a style to spread across the scene.

Has GPT Image 2.5 or Nano Banana Pro fixed logos?

They have largely fixed spelling. Rendering a word and reproducing a registered mark are different problems. The vendors’ own docs still say to spell brand names letter by letter, to state that geometry and layout must be preserved, and that fine details like clothing logos may not match.

Is it a legal problem to publish a slightly wrong version of my own logo?

It is a brand problem first: consistent use of the mark as registered is what protection rests on, and a stream of variants blurs it. Distorted versions of someone else’s mark are the real exposure; the Getty v Stability courts treated distorted watermarks in AI output as a trademark question.

What is the fix?

Never ask the model to draw the logo. Generate the scene without it, composite the real file on top with placement rules, lock colors and fonts as assets, keep the clean original for edits, and put a human gate before anything ships.

The bottom line

Image models have become very good at scenes and reasonably good at words. They are structurally incapable of being exact about a specific mark, because generation is reconstruction and a logo has one right answer. The vendors say so in their docs, the forums say so in their titles, and two courts have said a distorted mark in AI output is still a mark. Stop asking the model to draw it. Generate the scene, place the file, keep the original, check at 400 percent, and the most common failure in AI marketing images stops happening.

Marqeable runs your campaigns, answers every visitor, text, and email in seconds, and turns them into booked jobs and meetings - even at 9pm on a Saturday. We’re in private beta with a small early cohort. Get early access