I wanted to know something specific: if you give NB2 an impossible concept and tell it to sell it, does it just slap on some ad-looking text? Or does it actually think like an ad agency?

So I ran a test. Forty images. Ten variations. Two rounds. The concept: take things that can't physically be sliced open - clouds, ocean waves, flames, lightning bolts, the aurora borealis - render them as photorealistic cross-sections revealing their impossible internal structure, and frame each one as a luxury advertisement for a different product category.

What I found was bigger than I expected.

Five subjects that have no insides. Five ad categories that have very specific visual languages. One prompt structure holding it all together.

The pairings weren't random. Each subject maps metaphorically to its ad category. Cloud and fragrance - selling the invisible through visual metaphor. Ocean wave and premium spirits - layered depth, aged complexity. Flame and automotive - precision engineering, controlled power. Lightning and tech - invisible energy made tangible. Aurora and skincare - inner luminosity, revealed radiance.

The constant across every prompt: "Clean photorealistic cross-section of [subject]" - an anchoring phrase that keeps the cross-section concept from getting bulldozed by the ad-category descriptors.

Round 1 gave NB2 creative freedom. I specified the concept and the ad category. That's it. NB2 invented everything else - brand names, taglines, product placement, visual identity.

Round 2 put me in the director's chair. Same concepts, but I specified exact text, spatial placement, and layout. I wanted to know: can you art-direct NB2 like a creative director?

First prompt fired. Four images came back. Every single one rendered the cloud's interior as agate mineral banding - concentric opalescent rings, crystalline cavities, the aesthetic of a sliced geode. Not one went wispy or atmospheric. NB2 found the most structurally plausible interpretation of "what's inside a cloud" and committed.

But here's what stopped me. I never mentioned a perfume bottle. I said "luxury perfume advertisement" and NB2 generated bottles, brand names, taglines - four completely different fragrance houses, each with French or Latin naming conventions. "Cumulus D'Or." "Nuage d'Opale." "Aetheria Paris." One image even designed a cloud-shaped bottle cap.

Average score: 9.00 out of 10. The concept worked.

Same prompt structure, whiskey ad context. The wave resisted. Where clouds accepted a clean slice (they're solid-coded - they look like you could cut them), waves are fluid-coded. NB2 couldn't figure out how to bisect water. Three of four images drifted - transparency instead of cross-section, split-level photography, one image dropped the ad component entirely and just rendered a beautiful wave.

Only one image held: NB2 invented a gold-framed display window cut into the wave, like a gallery viewport into the ocean's interior. That was the creative workaround for a subject that refused to be sliced.

Average dropped to 8.57. My lowest score. But it told me something important.

Two words: "surgically bisected."

I added it to the remaining three prompts. The phrase does two things - reinforces the physical cut and adds a clinical, specimen-like tone that makes the slicing feel intentional rather than vague.

Result: bisection rate jumped from 63% to 92%. Average score climbed +0.32 points. It held even for lightning - the most abstract subject in the set, a thing with no surface, no edges, no stable form.

That's a replicable prompting technique. If you're trying to render a cross-section of something ephemeral, "surgically bisected" is the phrase that makes it stick.

Here's where it got weird. And I mean that in the best way.

I expected NB2 to change surface-level things across ad categories. Different colors, different composition, different text styling. That happened. But it also changed what was inside each impossible cross-section.

Cloud interiors for a fragrance ad? Geological. Opalescent mineral banding. Organic and precious.

Flame interiors for an automotive ad? Engineered. Hexagonal lattices, honeycomb grids, precision channels. Like the flame was designed by a mechanical engineer.

Lightning interiors for a tech ad? Digital. Fiber-optic cables, circuit board patterns, neural branching networks.

Aurora interiors for a skincare ad? Flowing mineral. Polished opal-like ribbons. Luminous and organic.

The ad category wasn't just styling the image. It was authoring the anatomy.

Five subjects. Five ad categories. Five completely different impossible biologies.

Three times across the test - independently, in different variations - NB2 did something I hadn't seen before. Instead of rendering a beauty shot of the cross-section, it switched to engineering cutaway diagram mode. Technical annotation lines. Labeled components. The impossible subject presented as a manufactured device with documented parts.

A flame with labels reading "Crystalline Heat Lattice" and "Calibrated Precision." A lightning bolt with components called "Data Plasma Conduit" and "Photonic Stabilizer."

NB2 invented science fiction spec sheets for natural phenomena. And presented them with the visual authority of an actual technical document.

Quick sidebar. We know from previous testing that white backgrounds carry roughly a 33% score penalty. Skincare ads demand white backgrounds. So what happens?

The penalty didn't show up. The beauty ad context overrode the known limitation. NB2 treated the white background as a design requirement and filled the space with marble surfaces, glass display stands, and prismatic light refractions.

The takeaway: ad-category context can neutralize known background penalties when the category demands that background.

Round 1 proved NB2 has ad intelligence. Round 2 asked: can you control it?

I rewrote all five prompts with specific instructions - exact brand names and taglines controlling what text appears, product placement like "positioned lower right on white marble," text positioning like "centered at bottom" and "upper left corner," and spatial composition like "positioned left of center" and "filling upper half."

The brands I created leaned into the absurdity of the concept. A perfume called Nimbus No. 9 with the tagline "Harvested at 30,000 Feet." A whiskey called The Drowned Barrel - Trench Reserve 21 with "Distilled at the Bottom of the Sea." A car brand called Pyre Motors with "0 to Inferno in 3.1 Seconds." A chipmaker called Zeus Semiconductor selling the Bolt X1 with "A Storm in Every Socket." A skincare line called Borealis Beaute selling Polar Luminance Serum with "Apply the Northern Lights."

Absurd origin stories delivered with straight-faced luxury ad cadence.

Sixty specified text elements across twenty images. Every brand name. Every tagline. Every product label. Zero errors. Zero garbled characters. Zero substitutions.

NB2 rendered "BOREALIS BEAUTE" with the accent on the E. It rendered "0 TO INFERNO IN 3.1 SECONDS" with the decimal point. It rendered "NIMBUS No. 9" four times identically.

And here's the thing - it didn't just follow instructions. It supplemented them. When I specified "NIMBUS No. 9" on the bottle, NB2 added "EAU DE PARFUM" underneath, because that's what goes on a fragrance bottle. When I specified "TRENCH RESERVE 21" on the whiskey, NB2 added "AGED 21 YEARS" because that's standard spirits label formatting. When I specified "50ml" on the skincare bottle, NB2 added the EU "e" mark - the European Union estimated quantity compliance symbol that's required on cosmetics sold in Europe.

NB2 knows EU cosmetic packaging regulations. I did not expect to type that sentence today.

Five for five. Every variation scored higher with art direction.

The biggest lift was the wave - Round 1's weakest variation became competitive. Spatial framing instructions ("filling upper half of frame," "placed lower center on barnacle-covered dark stone") gave NB2 a compositional structure that water couldn't dissolve. The prompt didn't just describe the image. It described the layout.

Variation C2 produced the highest-scoring single image in the test: 9.63. A flame cross-section with hexagonal lattice internals, centered on a carbon fiber engine mount, with "PYRE MOTORS" upper left, "COMBUSTION GT" lower right, "0 TO INFERNO IN 3.1 SECONDS" centered bottom. Perfect triangulated text layout. First perfect prompt-alignment score.

And NB2 designed a flame-shaped logo to go next to the brand name. Nobody asked for a logo.

But the image that actually made me pause was from the same variation. Image 2 showed a literal metal blade running vertically through the flame - the surgical instrument still embedded in the specimen. NB2 interpreted "surgically bisected" not as "show the inside" but as "show the tool that made the cut." Forty images in, and it still surprised me.

A pattern I'll use going forward: single material phrases in prompts function as storytelling devices. "Barnacle-covered dark stone" didn't just describe a surface - it made every whiskey decanter look like it had been pulled from the ocean floor. "Carbon fiber engine mount" didn't just place a flame - it made the fire feel engineered. "Marble surface" didn't just set a scene - it made the aurora slab feel precious.

One phrase. Massive narrative impact.

Forty images. Ten variations. Overall average: 9.15.

Highest variation: Aurora x Skincare, art-directed (9.43). Highest image: Pyre Motors, perfect layout (9.63). Art direction lift: +0.33 average across all five variations. Text accuracy: 60/60 (100%). "Surgically bisected" lift: +0.32 average, bisection rate 63% to 92%. Nine of the top ten images came from Round 2.

When you want NB2 to improvise the campaign:

When you want to control the layout:

Subjects: cloud, flame, ocean wave, lightning, aurora, tornado, sunbeam, fog bank, rainbow, smoke plume

Internal structures: layered opalescent structure, crystalline architecture, fiber-optic pathways, flowing ribbon layers, geological strata, branching networks

Ad categories: luxury fragrance, premium spirits, luxury automotive, technology brand, luxury skincare, haute couture, fine jewelry, premium timepiece

Surfaces that tell stories: barnacle-covered dark stone, carbon fiber engine mount, white marble, volcanic glass, driftwood, brushed titanium

NB2 doesn't just apply ad styling to images. It thinks in ad categories. It knows that fragrance demands French naming conventions and that spirits labels need age statements. It knows that automotive ads worship precision geometry and that skincare ads need dropper bottles, not pumps. It knows the EU compliance mark for cosmetic packaging.

And when you give it an impossible subject - the inside of a cloud, the anatomy of a lightning bolt - it doesn't just make something up. It makes something up that fits the industry.

That's not image generation. That's campaign intelligence.

Testing methodology: NB2 (via Adobe Firefly). All images scored using a weighted 5-dimension rubric. Minimum 4 generations per variation before drawing conclusions.

I'm an Adobe Firefly Ambassador documenting replicable patterns through systematic testing. Follow along - I publish new findings every week.

@GlennHasABeard | @AdobeFirefly | @NanoBanana