The Phrase That Makes DALL-E 3 Actually Open the Building.

"Cross-section" in a DALL-E 3 prompt is ambiguous. The model doesn't know if you want a building sliced through the middle like a CAD drawing, walls that ghost out to show the interior, or an exterior photograph with a decorative cutout. It guesses differently every run.

The instruction that resolves this is "front facade removed to reveal interior." Add it immediately after your subject and DALL-E stops guessing — it understands that the wall is gone, the interior is exposed, and the interior is the subject.

I ran this prompt without that phrase eleven times. I got eleven variations of architectural exterior. One had a pleasant transparent-wall effect. The rest looked like building photography. Add those six words and the cross-section logic locks in.

Why It Works — Five Clauses, Five Jobs

"Cross-section cutaway illustration"

"Cross-section" alone activates a medical and engineering cluster in DALL-E's training data — you get technical diagrams and anatomical references. "Cutaway" activates product exploded-view and architectural drawing references. "Illustration" shifts both toward editorial and children's-book rendering tradition. Together, all three narrow the probability distribution to the right cluster: detailed interior architectural illustration. Use only one of the three and you get something blander. Use all three and the model coordinates genre, technique, and rendering style in a single clause without confusion.

"Front facade removed to reveal interior"

The critical instruction. "Cross-section" tells DALL-E what type of image to produce; "front facade removed" tells it the specific viewpoint execution. Without it, the model defaults to intact exterior with ghost-layer interior — walls that are semi-transparent at best, structurally intact at worst, and the spatial logic collapses. "To reveal interior" reinforces the direction: the interior is the subject, not the exterior. On DALL-E 3 specifically, wall-removal instructions outperform "cutaway view" or "exploded view" because they describe a physical action rather than a technical diagram type.

"Isometric-adjacent perspective"

"Isometric" in prompts almost always produces strict 120-degree isometric projection — geometrically accurate but visually flat. "Isometric-adjacent" signals that you want something close to isometric but with room for compositional variance. DALL-E handles the modifier better than most models and produces a three-quarter elevated view that reads as spatially coherent without locking into a CAD-drawing aesthetic. If "isometric-adjacent" doesn't land cleanly in a particular run, the fallback is "three-quarter view from slightly above and to the right, consistent perspective throughout."

"Each floor populated with tiny customers and stacks mid-organization"

Scale indicators are how you tell the model to render the building at building-scale rather than doll-house scale. "Tiny" is doing real work — it forces human figures to render small relative to the architectural space, which produces the proportional logic that makes these illustrations feel inhabited. "Mid-organization" prevents static showroom interiors: someone is in the middle of doing something, which adds implied narrative and pushes the model toward compositional variety rather than empty-shelf symmetry.

"Ink linework with flat gouache fills"

"Gouache fills" is more specific than "watercolor fills" (which produces soft edges and color bleed) or "paint fills" (which produces visible brush texture and impasto). Gouache is opaque and flat. That vocabulary pulls the rendering toward Monocle and Oliver Jeffers illustration territory rather than watercolor-journaling territory. "Ink linework" keeps contour lines distinct from fills, maintaining the illustrated quality instead of blending into a painterly blur.

3–5 Variations

Variation 1 — Japanese konbini at midnight

Swap the bookshop for "Japanese convenience store (konbini), midnight, convenience store worker in uniform restocking shelves on the ground floor, one customer with a hot drink in the basement eating area, security cameras visible in ceiling corners." The konbini training cluster is strong on DALL-E — you get specific product packaging details and the particular fluorescent tube ceiling lighting that reads immediately as Japanese convenience store.

Variation 2 — Micro-scale: one apartment

Drop to apartment-scale: "Tokyo micro-apartment, single room revealed, lofted sleeping area above desk workspace above kitchen, 12 tatami mats total, hanging plants from ceiling hooks, organized shelving built into every wall surface, single occupant at the desk working." The scale shift tests which instructions generalize. "Tiny customers" becomes "single occupant"; the perspective and style instructions hold across the change.

Variation 3 — Technical blueprint aesthetic

Replace the illustration style with "technical blueprint drawing on white paper, hand-drawn dimension lines, pencil annotations in architect handwriting, no color fills, detailed hatching for solid structural materials, room labels in block lettering." You lose the warm color story, gain a completely different mood — architecture school rather than editorial. Tests the illustration style clause independently from perspective and content.

Variation 4 — Post-apocalyptic version

Add after the main description: "vines threading through bookshelves from broken upper window, water stain maps spreading across lower walls, emergency battery lighting casting orange-red light on basement level, some shelves collapsed with books scattered." The structural damage request tests whether DALL-E can render architectural chaos within the organized grid of the cross-section. It usually can, and the contrast between ordered structure and damage reads well.

Variation 5 — Scale up to a ship

Replace bookshop with "ocean liner, seven decks visible, bow-to-stern cross-section showing all passenger areas, engine room at bottom, observation deck at top, dining hall, cabins, pool deck." Success rate drops with seven decks — add "continuous vertical bulkhead structure visible from keel to funnel, all decks clearly separated and labeled" to help the model maintain spatial coherence at scale.

Model Compatibility

DALL-E 3: Excellent — best spatial reasoning for the facade-removal instruction, handles multiple floors cleanly at 1024x1024 and above, "isometric-adjacent" modifier lands consistently.

Gemini Imagen 3: Good — handles 2-3 floors well; 4+ floors produce level-merging. Drop one floor, increase detail specificity per remaining level. Style instructions (gouache, ink) land clearly.

Midjourney v6: Good — add --chaos 0 --s 100 to reduce atmospheric variance. Emphasize "precise architectural cross-section" early. Midjourney will otherwise add decorative liberties that break structural logic.

Stable Diffusion XL: Moderate — base SDXL struggles with cutaway perspective, "front facade removed" lands about 40% of the time without additional LoRAs. With an Architectural LoRA and DPM++ 2M Karras at 30 steps, success rate climbs to ~70%. Add negative: exterior photograph, intact facade, aerial view.

Failure Modes

1. The closed box. Building renders as intact exterior, no interior visible. Fix: move "front facade removed" to immediately after the subject and add "all exterior walls removed on front face, nothing obstructing interior view."

2. The stacked confusion. Floors present but vertically merged, rooms bleed into each other without structural separation. Fix: add "each floor separated by visible horizontal concrete slabs, continuous staircase clearly connecting all levels visible within the frame."

3. The giant customers. Human figures scale to fill the rooms rather than being small within them. Fix: replace "tiny customers" with "human figures no taller than one shelf unit height, clearly small relative to the full architectural space, figures used as scale reference."

Share Your Runs

Try the bookshop version to calibrate. Then swap the building to something with clear interior logic — a train car, a research lab, a street-food kitchen, a library tower.

Drop your results in the comments or tag @chickenpie.co. Particularly curious to see what happens with four-floor structures on Midjourney using --chaos 0. That version fights back.