The test: one brief, three rough doodles
This is not a comparison of three subjects or three art styles. Every run asked for the same unbranded amber perfume photograph, with the same stone materials, warm light, and restrained colors. The variable was the drawing: centered, diagonal, then deliberately open on the left.
Here is the exact prompt used for all three generations. It describes the finish in detail but never tells the model where to place the bottle, arch, platform, or foliage. Those decisions have to come from the attached drawing.
Use case: product-mockup
Asset type: editorial before-and-after experiment for a web guide
Primary request: Use the attached rough blue doodle as the composition reference.
Turn it into a high-end, original, unbranded perfume campaign photograph.
Input image: Image 1 is a layout reference, not an edit target.
Scene/backdrop: pale limestone studio set with subtle botanical elements only
where the doodle suggests them.
Subject: one translucent amber-glass perfume bottle with a simple blank front.
Style/medium: photorealistic luxury product photography with convincing glass,
stone, and natural texture.
Lighting/mood: warm directional morning light from the upper left, restrained
cream and olive palette, elegant but not glossy or synthetic.
Constraints: preserve the bottle's relative position and scale; interpret the
doodle's large curves, platform, diagonal, and plant-like strokes as set
architecture, shadows, or foliage; keep untouched areas genuinely uncluttered.
Avoid: words, labels, logos, brands, people, hands, extra bottles, frames, split
screens, or drawn blue lines in the result.Attempt A: the safe, centered composition
The first doodle is almost comically bare: a bottle in the middle, a large arch behind it, an oval underneath, and two forked marks near the edges. There is no perspective grid, glass rendering, shadow, or carefully drawn leaf. It is enough to say, “formal, symmetrical, product first.”
The transformation is the surprise that makes this workflow worth trying. The uneven rectangle became faceted amber glass. The oval became a heavy limestone stage. The two awkward forks turned into olive branches, and the arch became a sculptural opening with believable thickness and texture. The model invented nearly every surface detail, but the visual hierarchy still reads like the doodle.
As a product hero, A works. The bottle is unmistakably important and the symmetry feels expensive. Its weakness is just as clear: there is no obvious place for a headline. Adding “leave room for copy” to the prompt might help, but the drawing itself keeps pulling attention toward the center.
Attempt B: more energy, but the wrong kind of space
For B, we moved the bottle down and right, added a low rectangular platform, and swept one large diagonal stroke through the frame. Short offshoots made that stroke vaguely botanical. A rough circle sat in the upper-right corner. The left side was no longer empty; it contained a big compositional gesture.
We did not rewrite the prompt to explain any of that. We sent the same words again.
The output is arguably more dramatic than A. The bottle stayed on the right. The platform became a stone slab. The diagonal stroke became a rising limestone wall, its little offshoots expanded into olive foliage, and the circle became a pale stone disc. The model did not trace the marks; it found plausible photographic objects that could occupy the same areas.
But we would reject B for the intended use. It is a beautiful image, yet the diagonal wall and plants make the left half too active for a clean headline. That is not a rendering failure. It is a design failure created by our own sketch. A longer prompt would be an indirect way to fight the instruction already embedded in the drawing.
Attempt C: fix the layout with the eraser
The third sketch is a correction of B, not a fresh concept. We kept the bottle low on the right and retained the architectural curve, but removed the sweeping mark from the left. The only plant strokes now sit against the far-right edge. Most of the left half is literally untouched.
Then we sent the same prompt for a third time. No extra sentence about advertising space. No negative-prompt paragraph. No request to “make the composition more professional.”
This time, the model built the set on the right and left a broad, calm field of warm stone on the left. The bottle grew somewhat larger than the doodle suggested—a useful reminder that a sketch guides rather than locks geometry—but its placement, the arch, the plinth, the right-edge foliage, and the open left side all survived.
C is the strongest layout for an ad or landing-page hero because it solves a specific job. The important lesson is not that the third generation was magically better. We made the instruction clearer by deleting a line.
What changed, what survived, and what the model invented
| Run | What the doodle changed | What appeared in the photograph | Editorial verdict |
|---|---|---|---|
| A | Centered bottle, symmetrical arch, oval base, marks on both sides | Centered hero bottle, stone opening and round plinth, balanced olive branches | Polished and safe; weak copy space |
| B | Bottle moved right; large diagonal and branching marks crossed the frame | Right-weighted bottle; diagonal stone wall and dense foliage took over the left | Most dynamic; rejected for the intended headline layout |
| C | Removed left-side marks; kept bottle, arch, and foliage on the right | Right-side set with a broad, quiet field on the left | Best fit for an ad or page hero |
Across the three runs, the sketch consistently influenced four things: the bottle's general position, the direction of the largest structural line, where botanical detail accumulated, and where the frame remained quiet. It did not reliably lock exact proportions. It also did not specify whether a circle should become a sun, stone disc, or opening; the model chose an interpretation that matched the photographic brief.
That division of labor is more useful than asking whether the model “copied” the drawing. A rough sketch communicates relationships. The prompt supplies identity, material, light, and constraints. The generator negotiates between them and fills in missing detail.
Redraw the sketch or rewrite the prompt?
When a result misses, diagnose the layer before adding more prose.
| The problem you see | Best next move | Why |
|---|---|---|
| The subject is in the wrong place, too large, or competing with another shape | Redraw the large shapes | The error is spatial; show the new relationship directly |
| A mark was interpreted as the wrong object | Add one plain label in the prompt | “The oval is a stone plinth” resolves meaning without rebuilding the layout |
| The composition works but the image feels generic | Change material, light, palette, or camera language | The sketch has finished its job; art direction belongs in words |
| One generated detail is wrong | Edit that detail and state what must stay unchanged | A local correction risks less compositional drift than a full regeneration |
The B-to-C change is the first row in practice. We wanted calm negative space, so we removed the mark occupying that space. OpenAI's current image-generation guidance similarly recommends small, targeted revisions and explicit instructions about what should change and what should stay fixed.
Change only the bottle cap from clear amber glass to brushed brass.
Keep the bottle shape, position, plinth, arch, foliage, camera, crop, lighting,
shadows, and empty left side unchanged. Do not add text, a label, or a logo.How to use ChatGPT Sketch on mobile
OpenAI's current mobile instructions describe this flow:
- Type
@in the ChatGPT message box and choose Sketch. - Draw only the large objects and their relationship. Use undo or erase when a mark no longer helps.
- Tap the checkmark to confirm the drawing.
- Add a written instruction that identifies ambiguous shapes and describes the desired finish.
- Send the message, inspect the generated image, and continue in the same conversation for targeted edits.
Feature placement and availability can vary by plan, workspace, device, and rollout. If Sketch is not visible in your composer, you can still apply the method by attaching a simple drawing as a reference image and explaining what the shapes represent.
Try the three-thumbnail test yourself
Do not begin by polishing one drawing. Give yourself two minutes and make three disposable versions of the same idea:
- Safe: put the subject in the center and use a balanced background.
- Different: push the subject toward an edge and add one strong diagonal or foreground shape.
- Useful: decide where text, a face, or another important element must go, then leave that area completely unmarked.
Run the same prompt against all three. Judge the large structure before judging texture or beauty. Which image gives the eye a clear path? Which one protects the space required by the final use? Which invented details help, and which came from marks you should erase?
This test is fast because none of the sketches deserves protection. If the first idea is only a few imperfect lines, discarding it does not feel like losing work.
Where a rough sketch stops being enough
Sketch is excellent for composition exploration, but it is not a production specification. Do not use a doodle as a substitute for an exact logo, packaging dieline, measured interior plan, engineering drawing, regulated label, or approved legal copy. The perfume images above are direction-setting concepts. A real campaign would still need the actual product geometry, brand assets, typography, retouching, and stakeholder approval.
The right expectation is wonderfully modest: you are not drawing the final image. You are giving the model something that words handle badly—a map of where things belong. Draw quickly, try more than once, and let the differences between the results tell you which idea is worth refining.
