Ten attempts across Porto, Singapore, and Marrakech produced the six examples kept for publication. The opening Porto pair exposed a specific failure, so we revised the workflow and produced initial Singapore and Marrakech pairs. A later evidence pass found the first Singapore flower too generic for the botanical wording in our record. Four new Singapore candidates followed, with two retained. That sequence matters because an attractive example alone cannot show whether the instructions or the review were accurate.
Define the editorial claim
Start by writing one sentence about the asset you are producing. A tourism-board cover, a personal trip recap, and an imaginary city mood board have different evidence needs. The tests here use this claim:
An original mixed-media editorial page representing one real destination through a small set of sourced, visible anchors.
That sentence rules out several shortcuts. The picture is not a documentary photograph. It does not claim that the depicted objects occupied one scene at one time. It does claim that the named items belong in the destination vocabulary and that the title identifies the intended place.
If the assignment is fictional, label it fictional. If it promotes a real venue, product, event, or itinerary, confirm the details with the owner. Visual style does not lower the need for factual review.
Build a source ledger, not a souvenir list
Create a small table with four columns: candidate element, authority, visible cues, and editorial reason. Prefer the institution closest to the fact. In our Singapore packet, Gardens by the Bay supported the form of the Supertrees; NParks supplied the orchid's flower-part and color description plus the otter's body and coat cues; the Land Transport Authority covered MRT; the National Heritage Board supplied tile context; and the tourism authority supported the breakfast set.
| Candidate | Evidence question | Image question | Decision |
|---|---|---|---|
| Landmark | Does an official or primary source identify its form and location? | Which two features distinguish it from a generic tower or bridge? | Usually keep one as the hero |
| Food | Is the association city-specific, regional, or only national? | Can a small portion remain identifiable? | Keep, qualify, or leave out |
| Plant or animal | Is it genuinely tied to the place, or merely possible there? | Which visible traits can survive a stylized illustration? | Describe reviewable morphology or omit |
| Transport | Is it locally characteristic and current or historic in the intended context? | Could the model confuse it with another vehicle? | Describe body shape and prohibit doubles |
| Craft object | Does a museum or heritage source provide context? | Can we request a new motif rather than replicate a protected design? | Use an original fragment |
A source ledger is also an editing tool. Two items may both be true but communicate the same idea. Drop one. Another may be correct at country level yet misleading as the centerpiece of a city. Demote or remove it. The goal is not to pack trivia into a page; it is to create a legible hierarchy with defensible associations.
Turn names into reviewable geometry
Models receive words, while reviewers see shapes. Translate proper nouns into visual cues before generation. Our Porto entry described two stacked decks carried by one broad iron arch above the river. The Marrakech hero specified a single square-profile Koutoubia minaret behind low rose-ochre medina walls. The Singapore description called for tall planted structures with broad branching crowns. Its corrected orchid entry also specified unequal pale sepals and petals, a bright pink-purple angel-shaped lip, an orange center, and spots rather than relying on the species name alone.
These descriptions do two jobs. They help generation, and they create a pass criterion. If the Marrakech output shows two minarets or a rounded tower, an editor can reject it without debating whether the mood feels Moroccan enough.
Choose exclusions from likely model associations
A generic negative phrase is rarely enough. Write down what could plausibly appear because it is adjacent, famous, or stereotypical. This is not a list of everything in the world. It is a short forecast of the model's most likely detours.
For Singapore we prohibited the Merlion, national flag, airport waterfall, lotus-shaped building, temple, colonial figure, another food set, another animal, a second standalone flower specimen, another train, and unnamed skyline landmarks. The wording explicitly permits green planting on the Supertrees and the flat floral motif printed on the one tile. For Marrakech we excluded camels, dunes, snakes, fez hats, the blue-city palette, palaces, beaches, flags, carpet-shop imagery, people, costume, extra food, extra lanterns, and extra minarets.
Run the base prompt twice before you trust it
One output can hide a systematic ambiguity. We submitted the Porto packet in two separate calls. Both pictures delivered a convincing handmade composition, the double-arched bridge setting, Ribeira houses, francesinha, azulejo, a detached yellow tram, and correctly spelled PORTO and PORTUGAL.
They also both added a second yellow rail vehicle on top of the bridge. The official destination information explains why the association is plausible: a metro line uses the upper deck. But the approved inventory requested only one tram fragment. We marked both runs as failures on closed-set compliance.
The repair was not to erase the tram or deny the bridge's real use. We added a destination-aware exclusion field that can distinguish “retain one separate historic tram” from “do not add any train or metro vehicle on the bridge.” This is more actionable than expanding the prompt with general adjectives.
Decide which categories deserve space
Keep the complete visual inventory between five and seven subjects, counting the landmark as one; do not fill every category automatically. Each item should add a different scale or meaning: built form, horizon, living detail, everyday life, movement, material culture, or food. If two candidates compete for the same visual role, keep the clearer one and mark the other field OMIT. Count each separate cultural object individually, while treating a natural breakfast or drink grouping as one semantic element.
Our Marrakech card intentionally left the animal and transport slots empty. The remaining set created useful scale changes: a tall architectural silhouette, a distant mountain strip, a small botanical sprig, a tabletop hospitality grouping, one tile shard, and one pierced metal lamp.
Both results retained one main minaret, ochre walls, the mountain horizon, olive foliage, the covered dish and tea glass, geometric tile, lamp, and the permitted title lines. No item from the prepared Marrakech exclusion list appeared. The two outputs are not identical, which is expected; approval concerns the defined content contract rather than pixel repetition.
Use two review passes
Separate meaning from finish. In the first pass, hide your aesthetic preference and compare the image directly with the ledger. In the second, assess whether the chosen craft language supports the publication.
Pass one: destination integrity
- Confirm the title character by character and count every occurrence.
- Match the hero against its written silhouette cues.
- Locate each approved supporting motif and note substitutions.
- Scan specifically for every prepared exclusion.
- List additional semantic objects, even when they are locally accurate.
Pass two: page construction
- Check that one fragment clearly leads the composition.
- Confirm that the warm base sheet remains visible as designed space.
- Look for torn edges, translucent pigment, imperfect ink, fiber, and restrained shadow.
- Reject glossy rendering, photographic cutouts, uniform sticker outlines, and interface-like grids.
- Inspect the final display size; a tiny correct object that cannot be recognized is not useful.
Understand what the corrected runs did not prove
The retained Singapore pair contained the selected landmark, waterfront context, breakfast, MRT, tile, and single title word. The orchids show the supplied unequal pale flower parts, bright pink-purple lips, orange centers, and spots; the otters show the supplied long, low, sleek profiles with lighter underparts. The prepared exclusions stayed absent. Those are meaningful manual observations, but they are morphology checks against the prompt rather than taxonomic identification or local expert approval.
Neither retained Singapore image demonstrated a literal seven-object scene. The model expanded tropical leaves and greenery as part of the setting. The Marrakech pair likewise used small background vegetation beyond the six semantic anchors. We therefore record literal count as false while allowing the images to pass the bounded semantic and visible-cue checklist.
This distinction prevents a common reporting error. “All selected subjects are present” is supported. “The model rendered exactly seven objects” is not. Closed-set product catalogs, safety diagrams, museum records, and regulated illustrations need a more deterministic assembly process.
Escalate review according to the stakes
A personal mood page may need only the creator's check. A destination publisher should add an editor who knows the place. A commercial campaign should involve the commissioning organization and its brand or legal review. Depictions of living traditions, religious settings, clothing, Indigenous culture, contested geography, or people deserve informed local review before publication.
Also separate factual suitability from rights. Writing an original prompt and generating a new composition does not automatically clear trademarks, venue restrictions, likenesses, or the terms of the generation service. Keep the source ledger, prompt, output file, model information when available, review notes, and final approval together.
Keep an evidence packet
For every accepted image, retain:
- the assignment and intended use;
- the source links behind each selected motif;
- the exact filled generation request;
- the original output rather than a social-media screenshot;
- dimensions, file size, and a checksum;
- the semantic and presentation review decisions;
- the reviewer, date, known limitations, and approval status.
Our machine-readable test record contains those technical fields for all six runs and preserves the exact Porto, Singapore, and Marrakech requests. The reusable Handmade Travel Scrapbook Collage Prompt packages the revised structure for a new fact card. It remains Structure Checked because bracketed templates are not identical to their filled test cases.
