A sticker pack is not one image. It is one character, eight or nine expressions, and a promise that every one of them is the same creature. Most first attempts fail at the promise, not at the drawing.
This guide covers the order of work: lock the character, build the set, ask for transparency, then check that each expression still reads at chat size.
Why a pack is a consistency problem, not a drawing problem
Eight images that each look fine on their own can still fail as a pack.
- The reader sees them side by side in a chat list, small, in one glance
- What must stay identical: silhouette, proportions, palette, line weight, eye shape
- What is allowed to change: pose, expression, props, and one accent colour
- Writing that boundary down before generating is most of the work
Get the boundary wrong and you get eight good drawings of eight different characters.
Write the brief as a set, not a picture
Six lines, this time describing the whole set.
- Character — species or archetype, plus two or three features that must survive every image
- Count and layout — “eight expressions in a 3×3 grid”, or one sticker per generation
- Expression list — name each one, in order; the model cannot guess your set
- Framing — same crop, same distance, same head size in every cell
- Line and palette — stroke weight, fill, and how many colours total
- Exclusions — no background, no shadow cast on the background, no text you did not list
The expression list is the line people skip, and it is the one that decides whether the pack feels designed or random.
Lock the character before you make the set
- Generate one clean portrait first, on a plain background, and keep it
- Feed that image back as a reference for the next expression instead of re-describing the character
- OpenAI describes the improvement directly: the model is better at working from reference photos, and distinctive features are more likely to carry through
- Multi-turn editing is documented to hold earlier changes consistent, which is what lets you build a set in one conversation
- The general technique is the one we cover in holding one subject across a set
The first image is the asset. Everything after it is a variation on a reference, not a new character.
Ask for transparency out loud
- Transparency is a documented instruction: the help centre states that images can follow instructions to make the background transparent
- The release notes add that the model handles more complex layouts including transparent backgrounds
- Put it in the prompt in words, and ask for a clean cut-out edge at the same time
- No edge-quality or delivery-size figure is published, so check the result at full size rather than assuming
- When the character has to sit on top of another image, see combining images in GPT Image 2.5
Transparency is a request, not a mode. If the prompt is silent, expect a background.
Keep each expression readable at chat size
Design for roughly 100 pixels, not for the poster version.
- Keep the head large in frame and the pose unambiguous
- Drop fine detail that disappears when scaled: thin props, tiny patterns, background objects
- One accent colour change per expression is usually enough to tell them apart
- If a sticker carries a word, keep it to two or three characters — see text rendering in GPT Image 2.5
If an expression only works at full size, it is not a sticker.
Two worked examples: a nine-cell pack and a single chat sticker
Both briefs use the six lines above. Run them as written, then change one line at a time.
- Nine-cell pack (3×3 grid). Prompt: “A 3×3 grid of nine stickers showing the same cartoon character, a round orange fox with big ears and a cream muzzle. Same head size and same crop in every cell, consistent thick black outline, flat colour, no shading. Expressions in reading order: happy, laughing, crying, angry, sleepy, surprised, thinking, waving, love. One small accent change per cell only. Transparent background, no drop shadow, no text, no borders. Square 1:1 image.” Check the result in this order: are all nine cells present, is the muzzle and ear shape identical across them, and are the expressions the ones you listed rather than generic smiles.
- Single chat sticker. Prompt: “One sticker of the same round orange fox with big ears and a cream muzzle, from the reference image. Front-facing, head and shoulders, thick black outline, flat colour. Expression: laughing with eyes closed. Transparent background, no shadow, no text. Square 1:1 image with the character centred and generous empty margin around it.” Check that the silhouette matches the reference exactly, that the edges are clean against both a dark and a light surface, and that the margin is even on all four sides.
The grid is the cheap way to test consistency; the single sticker is the one you actually ship.
Fix one expression without redrawing the pack
Regenerating the whole pack to fix one mouth is how a set loses its consistency.
- Comments can be placed directly on an image for more focused editing, the documented way to point at one expression
- The second documented route is to describe the change in the conversation panel without the selection tool
- OpenAI states the limit plainly: highlights are not always precise and edits may extend beyond the area you selected, so re-check the neighbours
- This is what makes one-expression corrections worth attempting — the workflow is in comment editing in GPT Image 2.5
Where sticker packs still break down
Plan for these, because no prompt removes them.
- Background residue: a cut-out edge often keeps a faint halo; check on a dark and on a light surface
- Drifting proportions: compare the head-to-body ratio across the set, not within one image
- Hands and props: the least predictable part of any generated character
- Exact brand lettering: generated wordmarks are reconstructions, not the real asset
- Small type: below roughly caption size it is a coin toss, whatever the release notes improved
The honest workflow is two passes: brief and generate the set, then repair the cells that drifted. Control details are in the GPT Image 2.5 overview.
Frequently asked questions
Can GPT Image 2.5 make a transparent background for stickers?
Yes. The help centre documents that images can follow instructions to make the background transparent, and the release notes state the model handles more complex layouts including transparent backgrounds. Ask for it in the prompt, then verify the edges at full size.
How do I keep a whole set looking like one character?
Generate one reference portrait first and feed it back into every later generation, changing only the expression line. The release notes state the model preserves subjects from reference images better and that distinctive features are more likely to carry through.
Should I generate the pack as a grid or one sticker at a time?
A grid is the fast consistency test and lets you compare expressions side by side. A single sticker gives you a larger, cleaner result per generation. Many people brief the grid first, then re-render favourites individually.
Can I fix one expression without regenerating the pack?
That is what comments on the image are for. Mark the cell, describe the change, then check the cells around it — OpenAI states edits may extend beyond the selection.