A logo is the one image you make that has to survive being almost invisible.
This guide covers the brief: what to pin, and how to check the mark.
What a logo brief must pin down
A logo prompt is a brief with the adjectives removed. Six lines do the work.
The six lines to fill in
Start from this skeleton and replace the parts with your own values.
Job: design a logo, flat vector look, transparent background.
Mark: geometric rounded monogram of the letters N and A, one connected shape.
Shape language: circular, rounded terminals, symmetric, uniform thick strokes.
Negative space: the counter between the two letters reads as a leaf.
Colour: flat solid colours only, two colours, no gradients, no bevel, no drop shadow, no 3D.
Lettering: nothing inside the mark; the name North Atlas is set separately.
Frame: one candidate, centred, generous margins, 1:1.The first line names the job, the last fixes the frame; everything between is a constraint.
- Mark type — monogram, abstract mark, pictorial mark, or mark plus name
- Shape language — geometric or rounded, circular or angular, symmetric or not
- Negative space — what appears inside the counter, or nothing
- Colour count — two or three flat solids, plus the exclusions
Lettering is the least reliable line and OpenAI publishes no accuracy figure, so keep a Chinese brand name out of the mark: rendering Chinese text in GPT Image 2.5.
Shape language and negative space
Shape language keeps a mark from drifting into decoration: geometric, rounded, circular or abstract.
Writing negative space
Negative space breaks first, so write it as a second shape and check it in every candidate.
- Name the primitives: circular monogram, rounded geometric, one-line abstract mark
- Fix the stroke: uniform thickness, no hairlines, no tapered flourishes
- State symmetry explicitly; asymmetry is the default
When the shape is easier to draw than to describe, draw it: Sketch is the documented way to hand the model a drawing as a reference, entered by typing @Sketch in ChatGPT. See drawing with Sketch in GPT Image 2.5.
Flat colour limits and transparency
Two flat colours is a constraint, not a style: write flat solid colours, no gradients, no bevel, no drop shadow, no 3D.
- Ask for transparency the way OpenAI documents it: make the background transparent
- Check it is real transparency, not white standing in for it, on a dark ground
- Flat fills survive scaling; gradients and soft shadows turn into banding
OpenAI documents the phrase make the background transparent, and says the model handles more complex layouts including transparent backgrounds. Nothing further is published, so fake transparency is something you catch by looking.
Several candidates from one image
Ask for a grid of the same mark and you compare candidates side by side instead of one at a time.
- Ask for four or six variants in one image, in a grid, with identical margins
- Vary one variable per cell: shape, negative space, stroke weight, colour pair
- Score across, not up and down: at 16 px, in one colour, reversed out
The grid is a prompt technique, not a documented feature: OpenAI publishes no variant count per request.
Checking the mark at 16 px
A logo is the one image that has to work when it is almost invisible.
| Size | What it tests | Pass condition |
|---|---|---|
| 16 px | Silhouette | You can name the shape |
| 32 px | Letterforms | The initials read without guessing |
| One colour | Structure | It holds with no palette help |
| Reversed out | Contrast | It stays legible on dark |
A mark untested at 16 px falls apart, unlike an illustration that only shrinks. OpenAI publishes no maximum resolution for GPT Image 2.5, so scale it yourself.
From a bitmap to a scalable file
What comes back is a bitmap, whatever it looks like.
- Use the bitmap as the drawing, not as the production file
- Retrace it in vector software, or run an automatic trace and clean the curves
- Check the retrace at 16 px again, because tracing thickens thin joins
Generate so the retrace is easy: no hairline strokes, no soft shadows, clean edges. Other controls are in the GPT Image 2.5 overview.
Worked examples: monogram and lockup
Two briefs you can run as written, one line at a time.
Case one: the abstract monogram
A two-letter monogram with a negative-space reading, two flat colours, transparent background.
Design a logo, flat vector look, transparent background.
Mark: circular monogram of N and A, one connected geometric shape, rounded terminals.
Negative space: the counter between the two letters reads as a leaf.
Colour: flat solid colours only, two colours, no gradients, no bevel, no drop shadow, no 3D.
Frame: one candidate, centred, generous margins, 1:1, no text inside the mark.Check in this order:
- Shrink to 32 px: do the two letters still read?
- Recolour to one flat colour on dark: does the silhouette hold?
- Look at the counter last, where negative space has usually closed up
Case two: mark with a brand name
A pictorial mark with the brand name outside it, so the name can become type later.
Design a logo mark plus the brand name North Atlas beneath it in a clean sans serif.
Mark: one abstract geometric shape, rounded, one flat colour, thick uniform strokes.
Name: wide letter spacing, straight baseline, generous gap between mark and name.
Colour: two flat solid colours, no gradients, no bevel, no drop shadow, no 3D.
Background: transparent. Frame: 1:1, generous margins, nothing else in the image.Check in this order:
- Shrink the mark to 32 px alone, then the name at its smallest print size
- Reverse the lockup out on dark and re-check the spacing
- Count the letters: one letter per pass beats a full re-run
If the name is Chinese, keep mark and name in separate images until both hold up; combining them is in combining images in GPT Image 2.5.
Logo design FAQ
Four questions that come up before a mark ships.
- Can GPT Image 2.5 make a transparent logo? Making the background transparent is the documented instruction; re-check the file on a dark ground.
- How many variants come back per request? That figure is not published, so the grid is a comparison technique, not a count.
- Can I register the result as a trademark? OpenAI publishes nothing about registrability, so treat a generated mark as a starting point and do a clearance search first.
- How do I fix one letter? Use comment editing in GPT Image 2.5, then re-check the letters beside it, since edits can extend past the selection.