Begin with an authorized, reviewable source
We used two real photographs from Unsplash. The first is a 1536 × 2304 Rotterdam photo by Alexander Psiuk with the Erasmus Bridge, skyline, river, and low sunset. The second is an 1800 × 1200 blue-hour winter-cabin photo by Himmel S from Levi, Finland. Both source pages identify the photographer and make the image available under the Unsplash License. We recorded the page URLs, license URL, dimensions, file sizes, and SHA-256 hashes before editing.
A travel photo can reveal more than a skyline. Before uploading your own image, zoom in and check for faces, home addresses, license plates, tickets, hotel keys, screens, documents, private interiors, and bystanders. Owning the file does not automatically resolve privacy, publicity, trademark, artwork, or commercial-use questions. If you plan to use the finished image in a product, campaign, or item for sale, keep an approval record appropriate to that use.
Understand transformation versus replacement
A text-to-image request such as “paint a beautiful Rotterdam skyline” generates a new interpretation. A reference edit begins with a specific photograph and asks the model to change selected visual properties. The difference matters. This workflow promises a stylized derivative that remains recognizably tied to the source—not a new city picture with similar landmarks.
| Request | What it controls | What it cannot prove |
|---|---|---|
| Generate a watercolor of a named city | Destination, mood, medium, broad landmark choice | Continuity with your photograph |
| Edit this photo into watercolor | Source-guided composition plus requested medium | Exact geometry or pixels |
| Apply a deterministic filter | Every source pixel through a known transform | Convincing handmade interpretation |
| Work with an illustrator | Human-controlled accuracy and expressive choices | Instant, low-cost output |
If exact facade lines, property conditions, distances, or documentary evidence matter, stop here and use a deterministic or human-controlled method. Generative style transfer is appropriate when recognizability can be judged visually and small interpretive changes are acceptable.
Map the landmarks that define the photo
Do not start with a long style paragraph. First mark three to seven facts that make the frame identifiable. Good anchors are visible, localized, and comparable at full size.
| Rotterdam source anchor | Protected relationship | Failure signal |
|---|---|---|
| Erasmus Bridge pylon | Large, right side, leaning geometry, split upper shape | Moved to center, mirrored, straightened, or redesigned |
| Cable fan | Thin diagonals descend left from the pylon; two stronger cables extend right | Random suspension pattern or missing side |
| Bridge deck | Crosses the lower skyline and meets the pylon base | New arch, disconnected deck, or shifted horizon |
| Main tower group | One tall tower left of two slimmer towers | Order, count, or relative height changes materially |
| Sun and light | Low left, warm glow behind the skyline | Moved sunset, night scene, or unrelated fantasy lighting |
| River | Occupies the lower frame with a cool blue-gray surface | Removed water, new shoreline, or dominant invented boats |
The prompt can emphasize those relationships without asking for a photographic copy. Ink pressure, pigment edges, paper texture, wash density, and simplified windows can vary. The landmark placement, building arrangement, viewpoint, and location caption should not.
The Levi source needs a different anchor map: cabin on the right, steep roof and chimney, porch and warm window, large snow-laden pine in the left foreground, smaller trees around the building, a slope rising to the right, and open blue-lavender sky above. The gallery direction is allowed to simplify bark, snow tracks, branches, and cabin texture, but it should not move the shelter or swap the dominant foreground tree.
Separate what should change from what should stay
OpenAI's current GPT Image prompting guide recommends clearly stating what should change and what should stay the same. For style transfer, it also suggests describing the visual qualities you want and setting firm layout constraints to reduce unwanted changes. A well-structured prompt is easier to review and revise than one uninterrupted paragraph.
| Prompt section | How we handled the Rotterdam test |
|---|---|
| Edit target | The downloaded waterfront JPEG was the only reference; no style image or earlier derivative was attached |
| Location | Rotterdam and the Erasmus Bridge had to remain immediately recognizable from the same portrait viewpoint |
| Composition | The right-hand pylon and cable fan, bridge deck, building arrangement, sky, skyline, river, low-left sun, and general reflection direction had to stay in place |
| Allowed style changes | Fine gray linework, translucent pigment, imperfect dry edges, slight paper bleed, and blue, gray, gold, peach, and cream drawn from the source photo |
| Page layout | One compact image centered on warm paper with generous open space and no surrounding decoration |
| Lettering rule | The verified place name ROTTERDAM once below the picture; no secondary words, stamp, or mark |
| Exclusions | No new subject, decorative motif, brand, watermark, photoreal finish, oversized title, or redesigned structure |
The public test record contains the exact Rotterdam-specific prompt used for both repeat tests and the exact Levi prompt used for the gallery test. The Prompt page offers a shorter four-variable version adapted from them. That reusable wording has not been run exactly as published, which is why it carries a Structure Checked label instead of Behavior Tested.
The prompt asks for a “Japanese travel-journal feel” instead of naming a living artist. You can describe the medium, paper surface, linework, texture, wash density, colors, and composition directly; an artist's name is not needed to achieve the desired look.
Build a gallery print, not just a watercolor filter
The more finished Levi result did not come from adding more painterly adjectives. It came from a clearer hierarchy. The scene is compact, the paper occupies most of the page, related cool colors carry the atmosphere, the source's warm window becomes the single focal accent, and the caption is visibly subordinate to the art.
| Layer | Rotterdam sketch mode | Levi gallery mode |
|---|---|---|
| Painted field | Large and detailed, with the landmark dominating the page | Roughly half the page width with more uninterrupted paper around it |
| Linework | Fine ink helps describe bridge cables and architecture | Minimal linework lets translucent winter washes carry the image |
| Color | Blue-gray, gold, peach, and cream follow the sunset source | Blue and lavender form one family; the cabin light is the restrained warm accent |
| Caption | One location label | Three small lines: mood, place, and year |
| Visual promise | Recognizable travel-journal illustration | Quiet stationery or gallery-print composition |
Place a relatively small version of the painted scene in the upper-middle of a
warm-ivory page. Aim for roughly 45–55% of the page width and no more than half
its height. Keep uninterrupted paper on every side and clear space below.
Reduce the source palette to a restrained family of related cool tones. Preserve
one natural warm-light accent when the source contains one. Use translucent
washes, visible paper grain, naturally imperfect edges, and minimal linework.
Do not add a frame, border, stamp, tape, prop, logo, or poster decoration.Image models may treat percentages as suggestions rather than measurements. Use them to communicate hierarchy, then judge the complete page at its real display size. If the image fills the page and the cream field looks like a narrow border, the art-print direction failed even if the watercolor texture is attractive.
Compare repeat runs to see what changes
We used the same detailed prompt twice with the unchanged source photo. The results are distinct files: test A has SHA-256 1bb65e64e3f6…, and test B has SHA-256 cdac2b562403…. Different hashes confirm that the files are not identical, but only a full-size visual comparison can show whether they are successful.
| Criterion | Run A | Run B |
|---|---|---|
| Location recognizability | Ready for final human review | Ready for final human review |
| Bridge pylon, side, and cable rhythm | Broadly preserved | Broadly preserved |
| Overall building arrangement | Broadly preserved; facade details simplified | Broadly preserved; small shapes vary |
| Skyline and river proportions | Close, although the painted area leaves slightly less sky | Close, with a denser painted area |
| Cloud and contrail shapes | Interpreted rather than exact | Interpreted differently |
| Brushwork in the water and reflections | Based on the source but rearranged | Based on the source and visibly denser |
| Caption | ROTTERDAM, spelled correctly and shown once | ROTTERDAM, spelled correctly and shown once |
| Paper and layout | Pass: ivory texture and open margins | Pass: ivory texture and open margins |
The prompt guides the overall style and composition more reliably than the smallest details. Both results are useful, but they do not prove that future runs will spell every location correctly or preserve a more complex skyline. The separate Levi output has SHA-256 970802b348ba…; it kept the cabin on the right, the roof, chimney, porch, warm window, foreground pine, slope, and blue-hour atmosphere while rendering all three requested caption lines correctly. Because gallery mode has only one run, it demonstrates a successful direction but says nothing about its repeatability.
Review the image in three passes
- Identity pass. Ignore the style at first and compare the landmark position, silhouette, bridge or shoreline, building arrangement, horizon, lighting, and crop with the source.
- Style pass. Check the ink variation, translucent washes, paper texture, source-derived colors, restrained detail, painted-field scale, and generous open space. Reject glossy vector art, polished 3D rendering, photorealism, a thin paper border, or a generic poster look.
- Publication pass. Read the caption character by character, look for extra text or logos, confirm that you have the rights needed for the intended use, preview the final crop where it will appear, write useful alternative text, and obtain final human approval.
View the entire image at once, then inspect it again at 100–200 percent. The full view shows whether the composition still feels like the place. The closer view reveals invented windows, broken cables, letter-like artifacts, unwanted borders, or paper textures that resemble compression damage.
Use a two-pass typography workflow for production
The Levi test rendered winter light, Levi, Finland, and 2026 correctly in one call. Treat that as one accepted result, not a promise that small lettering will always work. Generated text is reasonable for a personal experiment after careful review. When the asset will be printed, localized, reused as a template, or approved by a client, ask the image model to leave the caption area blank and add the words afterward as editable type.
| Use | Recommended lettering method | Required check |
|---|---|---|
| Personal one-off image | Generate the short caption with the artwork | Read every character; reject duplicates and invented marks |
| Website editorial image | Either method, but repeat the exact place in HTML alt text or caption | Do not make tiny pixels the only accessible label |
| Print, campaign, localization, or template | Generate blank caption space; typeset the approved words separately | Keep an editable master and export proof |
Decide whether to edit or start over
| Observed problem | Next action | Reason |
|---|---|---|
| Caption is misspelled; image otherwise passes | Make a caption-only edit and repeat the full preserve list | The failure is localized and easy to test |
| Wash is too saturated | Change only wash saturation and density | A single style variable can be isolated |
| Landmark moved, mirrored, or changed shape | Start again from the untouched source | The current image is no longer a reliable starting point |
| Many small buildings drift | Reduce the anchor list to the few location-defining masses or use a non-generative method | Exhaustive geometry may exceed the workflow's useful tolerance |
| Page feels cluttered | Remove only named decorations; restate no border, stamp, props, or extra text | Generic “minimal” wording may not block familiar poster motifs |
Prepare the image for its final use
For a social post, preview the actual platform crop and make sure the location caption remains legible after compression. For a web page, include the exact place name, credit, and context in the surrounding HTML; do not rely on tiny generated lettering as the only accessible label. For printing, check the intended print dimensions and resolution, order one proof, and see whether the warm paper background prints neutral or yellow on the selected paper stock.
Keep the untouched source, accepted result, exact test prompt, date, available tool or model information, source page, license page, and final approval together. “Made with AI” is not a complete record. The useful questions are which source was used, what changed, who reviewed the result, and where the finished image will appear.
Why this is a Prompt and a Learn guide
We publish this technique as both a ready-to-copy Prompt and a Learn guide. It does not need a separate Skill because the process involves one upload, one structured edit, and one visual review—no installation, special permissions, dependencies, or repeatable agent workflow. The Hand-Drawn Style Director Skill is better suited to directing a series of original illustrated scenes. For a more general approach to preserving source details during an edit, read How to Edit Images With ChatGPT Without Losing the Original Scene.
The watercolor travel-photo test record contains both source hashes, both exact scene-specific prompts, all three output hashes, image dimensions, review decisions, and evidence limits. OpenAI documents GPT Image 2 as supporting image editing and high-fidelity image inputs. However, the tool used for our tests did not reveal its underlying model, and we did not test Gemini or Grok, so we do not describe these images as direct or cross-model results.
