Every prompting guide tells you the same thing: be specific. Name what you want. Spell it out. I've preached this myself - my own data from Part 1 and Part 2 shows specific prompts outperform vague ones by about a point on a 10-point scale.

Then I tested gravity, and the rule broke.

I wanted to know if Firefly could handle selective physics - gravity working normally for some objects while failing for others. Not everything floating. Not a zero-gravity scene. The hard version: a woman eats breakfast while her food drifts upward off the plate, but her hair hangs normally and her coffee mug stays on the table until the coffee inside it decides to leave.

I wrote two versions of the same scene. One specific, one vague.

Specific:

Vague:

The specific version scored 7.72. The vague version scored 7.80.

That's not a big gap. But the direction matters - it's the opposite of what every other test I've run has shown. And when I looked at the images side by side, I understood why.

The specific prompt produced neat levitation. Toast and eggs arranged in tidy vertical columns. Coffee spiraling upward in a photogenic ribbon. The women had blank, commercial-model expressions. The whole thing looked like a food brand advertisement - which is exactly what it is. Adobe Stock is full of "flying food" photography where ingredients hover artfully for product shots. When I told Firefly exactly which foods to float and how, it mapped to that training data and gave me a very polished breakfast ad.

The vague prompt produced chaos. Mugs, spoons, toast, and scrambled eggs all floating at random heights and angles. One woman glanced sideways at the hovering objects with an expression that said "huh, that's happening." Another smiled. The floating was messier, less controlled, less photogenic - and more believable as actual gravity failure.

Real gravity failure would be messy. Neat hovering is a photography trick. Chaos is a physics anomaly.

This doesn't mean vague is always better. For material transformations, "constructed from" beats "made of" by a full point because precision matters when you're replacing one substance with another. But physics violations follow different rules. Chaos is more believable than control when reality is supposed to be breaking.

The specific-vs-vague comparison was interesting. The complexity spectrum was the revelation.

I tested five levels of gravity-violation complexity. Here's what happened:

The relationship is almost perfectly linear. More floating things = lower scores. And the gap between best and worst is enormous - nearly two full points.

The highest-scoring variation (8.75 average, 9.25 peak) was a scholar reading at a library desk while books floated off the shelves around her, pages fanning open mid-air.

Three things made this work that the kitchen couldn't match.

First, there's no "flying books" advertising category in Adobe Stock. Firefly couldn't take a training data shortcut - it had to construct the scene from scratch. Same principle I found in Part 1: when no shortcut exists, Firefly builds better impossible imagery.

Second, dark environments hide the seams. The deep shadows of a Rembrandt-lit library conceal the transition zones where floating meets grounded. In the bright kitchen, every detail is exposed and every physics error is visible. Darkness is forgiving.

Third, "pages gently fanning open" gave Firefly a physical behavior to render. The books aren't just hovering - they're doing something. The fanning pages suggest a force acting on them, which transforms static levitation into active physics. It's the gravity equivalent of "constructed from" - a specific physical instruction that gives Firefly something concrete to execute.

The second-highest-scoring variation (8.64 average) replaced floating books with floating furniture - an entire living room's worth of couches, cushions, and lamps hovering near the ceiling while a person stood on the bare floor below, looking up.

Floating books are magical. A floating couch is ominous. The difference is mass. Heavy objects overhead imply danger - that couch could fall. The viewer doesn't just observe the impossibility; they feel the threat of it. And the empty floor where furniture should be creates its own kind of wrongness - absence as evidence that something is deeply off.

Every person in these images looks up. I didn't need to prompt it explicitly - Firefly understood that a person standing under ceiling-level furniture would naturally crane their neck. And the floating furniture casts shadows on the walls, confirming the objects exist in the real scene's physics, not pasted on afterward.

Then I stripped the concept down to its absolute minimum. One mug of coffee on an office desk. The coffee pours upward, pooling on the ceiling. An office worker types at their computer three feet away, completely unbothered.

This scored 9.18 - the second-highest single image of the session. One impossible thing. Everything else aggressively mundane. Fluorescent office lighting. Cubicle dividers. A guy typing emails.

The coffee stream is beautiful - amber liquid rising in a thick column with visible surface tension, small droplets breaking free at the edges. And it pools against the ceiling exactly as liquid would if gravity reversed for this one object. Firefly reasoned about what upward-pooling liquid would actually look like when it hits a surface. That's not placement - that's physics simulation.

The mundanity of the setting is doing half the work. The more ordinary the environment, the more jarring the single violation becomes. This image is closer to a Magritte painting than to a VFX shot - it's not spectacle, it's wrongness.

At the other end of the spectrum, I tested maximum complexity: an outdoor farmers market with fruit rising from crates while vendors and shoppers walked past, undisturbed.

It scored 6.89 average - the session's lowest by a mile. Three things compounded.

Flat overcast lighting exposed everything equally. There were no shadows to hide in, no atmospheric depth to suggest mystery. The floating fruit looked placed, not anomalous.

Twelve people ignoring floating fruit isn't surreal - it's two photographs badly merged. One person calmly eating while food floats is a choice, a characterization. A crowd unanimously failing to notice the impossible is just unconvincing. My Part 1 research proved that human reaction sells impossibility. Human non-reaction at scale kills it.

And the scene was simply too complex. Multiple stalls, multiple people, multiple floating objects, open environment. Every additional element was another place for the illusion to break.

Five variations, twenty images, one clear pattern. Here's what I'd take forward:

Float fewer things. The complexity curve is nearly linear. One floating object in a normal room (8.58) beats a room full of floating chaos (7.76) by almost a point. The ideal is a clear rule - one type of thing defies gravity, everything else doesn't - rather than widespread anomaly.

Use atmospheric environments. Dark wood and warm lamp light (8.75) beats bright kitchen (7.76) beats flat outdoor (6.89). Shadows hide physics errors, mood amplifies mystery, and depth creates the visual space that impossible objects need to exist convincingly.

Give floating objects behavior. "Pages gently fanning open" outperformed static hovering. "Coffee pooling on ceiling" outperformed coffee simply floating. Objects that do something while defying gravity are more convincing than objects that merely hover.

Let chaos be chaotic. Don't over-specify the arrangement. Neat levitation reads as advertising. Messy floating reads as physics failure. For gravity work specifically, vague is at least as good as precise.

React or don't notice - but don't have crowds ignore. One person looking up at floating furniture: powerful. One person typing while coffee defies gravity three feet away: surreal. Twelve people walking past floating fruit without a glance: unconvincing.

20 images. 5 scenes. The less you break, the better it looks.

Testing methodology: All images generated in Adobe Firefly Image 5, single session. Each variation generated 4 images. Scored on a 5-dimension rubric: Visual Quality (30%), Prompt Alignment (25%), Consistency (15%), Uniqueness (15%), X Engagement Potential (15%). No cherry-picking - all images scored, averages reported.