How to Put Your Logo on AI-Generated Images (Never Ask the Model to Draw It)
The most common failed check in any visual brand audit is the logo: missing on some images, three different sizes on the rest, two different corners, and on a few, a version the model “helpfully” redrew with a letter that is almost right. It is also the cheapest check to fix, because the fix is a rule, not a skill.
The rule is in the title. Generate the image clean, then composite the real logo file on top. This post explains why the model cannot be trusted with it, where the logo goes and how big, when to skip it, what to do when you edit the image later, a five-minute manual method, and what automatic placement does when a tool handles it for you.
Why the model gets your logo wrong
An image model does not paste; it draws. Ask it to include your logo, even with the logo attached as a reference, and it produces its best drawing of what it saw: the letterforms lose their exact geometry, the proportions stretch to fit the composition, the color drifts toward whatever the scene’s palette suggests, and any fine detail (a thin rule, a small icon, a registered mark) turns to mush. The result looks like your logo from across the room and wrong up close, which is precisely the kind of tell that makes viewers suspect an image and then distrust everything around it.
Reference fidelity has improved a great deal (GPT Image 2.5 holds a product’s label through edits better than anything before it), and it still does not make the model a paste tool. A logo is not a subject to be preserved approximately. It is a file to be placed exactly. Treat it that way and the problem disappears.
Placement rules
These are the rules that hold across social tiles, blog covers, email heroes, and one-off graphics, and they are simple enough to write on a sticky note.
| Rule | Default | Why |
|---|---|---|
| Corner | Bottom-right | Reads last, in the natural scan direction; least likely to fight the subject |
| Size | 8 to 12% of the image’s short side | Legible on a phone, unobtrusive on a desktop; sizing by short side keeps it consistent across ratios |
| Margin | About half the logo’s height from each edge | Room to breathe; survives platform cropping |
| Version | The one-color or reversed version for busy or dark backgrounds; full-color on plain light backgrounds | Contrast beats color fidelity when the two conflict |
| Contrast pad | A subtle, rounded, semi-transparent pad behind the logo when the background is close in tone | The logo should never depend on luck for legibility |
| Consistency | Same corner and size for every image in a channel | The feed reads as one company; the audit passes |
The exceptions:
- The default corner is busy. A face, readable text, or fine detail in the bottom-right means the logo moves to the least busy corner, and the choice should be by busyness (edge density, contrast) rather than taste. Sky and flat color are quiet; crowds and typography are loud.
- The image is tiny. Below roughly 200 pixels on the short side, the logo is illegible; skip it.
- The frame already carries it. An email template with a logo header, a branded social frame, a slide master. A second logo is clutter.
- Every corner is busy. Skip rather than force it; a logo crushed into a crowded corner reads worse than none.
Logo, not watermark. A watermark is a deterrent, tiled or centered, meant to survive theft. A logo is a signature, small and in a corner, meant to build recognition. Most marketing images want the second. If you are protecting product photography, the watermark tools are a different job.
The rule that saves you later: keep the original
Once the logo is composited, the image has two versions: the clean one and the branded one. Keep both, always, because the branded one is now the wrong input for anything else.
If you run an AI edit on the branded version (remix, upscale, reframe, inpaint), the model treats the logo as part of the picture and does what it does to everything else: redraws it. The correct sequence for any later change is original in, edit, logo re-applied on the result. That is easy when a tool keeps both versions linked and tedious when they are two files with similar names in a downloads folder, which is why the reference-images guide and every other post in this series keeps returning to the same point: consistency is a workflow, not an effort.
The five-minute manual method
For anything you generated in ChatGPT, Midjourney, or Canva and need to brand by hand:
- Export the logo once, properly. A transparent PNG (or SVG) at high resolution, in each version you use: full color, one color dark, one color light. Do this once and never again.
- Open the generated image in an editor that runs in the browser and keeps the file on your device; the free image editor does, and it is built for exactly this step.
- Drop the logo in the bottom-right, size it to about a tenth of the short side, set the margin to half its height.
- Check the corner. If it fights the subject or sinks into the background, switch to the reversed version, add a soft pad, or move to a quieter corner.
- Export as a new file with a name that says it is the branded version, and keep the original next to it.
Five minutes the first time, one minute after that. For sixty images a quarter, that is an hour, which is the argument for the next section.
What automatic placement does
The manual method fails in exactly one way: someone has to remember to do it, on every image, under deadline. Automatic overlay removes the remembering. A well-built version does the following, and it is worth knowing what to look for in any tool that claims it:
- Places the real logo file, not a drawing of it, after generation, so the mark is pixel-exact every time.
- Picks the corner by busyness, measuring edge density and contrast in each corner and choosing the quietest, with a tie going to bottom-right, rather than always stamping the same spot regardless of what is there.
- Sizes by the short side, so a portrait tile and a landscape header carry the same visual weight of logo.
- Adds a contrast pad only when needed, by reading the tone of the chosen corner.
- Skips when it should: images too small to carry a legible mark, or corners that are all busy, with a warning rather than a bad result.
- Keeps the original, and re-applies the logo automatically after any AI edit, so the branded version is always derived from a clean source.
- Lets you toggle it per image, because some images belong in a frame that already has the logo.
That list is how Marqeable’s logo overlay works: generated images are branded automatically with your uploaded logo file, placed by corner busyness with automatic contrast padding and short-side sizing, skipped on tiny images, toggleable per image with a chip in the draft, and the clean original is kept so remix, upscale, reframe, and inpaint edits operate on it and the logo is re-applied to the result. It is a small feature, and it is the one that makes the most common audit failure stop depending on anyone remembering. We are in private beta with a small early cohort: get early access if you would like the sticky note to become a setting.
Frequently asked questions
Can I use the logo as a reference image so the model places it correctly?
Using the logo as a reference will influence the model’s drawing of it; it will not make the model paste it. You will get a better approximation, not an exact mark. Composite instead.
What about text under the logo, like a tagline or phone number?
Overlay it in your real font as live text, next to the logo, in the same composite step. Never let the model render it; the text-in-images guide explains why generated type is the fastest tell.
Should the logo be on email hero images?
Usually not, because the email template already carries the logo in the header. On social images that will be seen out of context, yes.
How big is too big?
Above about 15% of the short side, the logo starts competing with the subject and the image reads as an ad. Below about 6%, it disappears on a phone. Ten percent is a safe default.
The bottom line
The logo is the one part of a marketing image that must be exact, and a model that draws everything cannot be exact about anything. Generate clean, composite the real file, same corner and size every time, move it only when the corner is busy, pad it only when the contrast is poor, skip it when it cannot be read, and keep the original for every edit that comes later. Do that by hand in five minutes or let a tool do it in none, but do it on every image, because the logo is what turns a good image into yours.
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
