I've been telling people to use "constructed from" instead of "made of" in their Firefly prompts for months. The data was clear: at portrait scale, swapping those two words improved scores by 2-3 points on a 10-point scale. But I'd only tested it at one magnification. What if the finding was a fluke of portrait photography?
So I took it to macro. Extreme close-up. 100mm lens territory. And I found something I wasn't expecting.
The A/B Test at Macro Scale
Same subject. Same background. Same lens. Same everything. The only difference: two words.
Prompt A ("constructed from"):
Extreme macro photograph of butterfly wing constructed from
intricate brass clockwork gears and mechanisms, tiny oil
droplets visible between gear teeth, soft diffused lighting
from above with subtle blue gradient background, 100mm macro
lens, shallow depth of field, professional macro photography,
hyper-realistic, ultra-detailedPrompt B ("made of"):
Extreme macro photograph of butterfly wing made of brass
clockwork gears, oil droplets between gears, soft diffused
lighting from above with subtle blue gradient background,
100mm macro lens, shallow depth of field, professional macro
photography, hyper-realistic, ultra-detailed
Same subject. Same lens. Same lighting. The left wing is engineered from clockwork. The right wing is decorated with it. Two words made that difference.
The gap: 0.98 points. Almost exactly one full point on a 10-point scale.
But here's what makes the macro test more revealing than the portrait test: at macro magnification, you can actually see what's different. In the "constructed from" images, the wing veins became brass rod frameworks. The gears are structural - they form the wing surface. Translucent membrane fills the gaps between mechanical sections. It's engineered. In the "made of" images, the gears sit on a wing that's still clearly a butterfly wing. They're decorative. Pretty, but not committed.
"Constructed from" told Firefly to build a mechanism in the shape of a wing. "Made of" told Firefly to put gears on a wing. The oil droplets prove it - mechanical coherence on the left, morning dew on the right.
Even the droplets told different stories. "Constructed from" produced oil between gear teeth - mechanically logical. "Made of" produced morning dew on a wing surface - biologically logical. Firefly was so tuned into the difference that it changed the type of liquid to match.
The finding holds. "Constructed from" vs. "made of" is now validated at portrait scale, architecture scale, and macro scale. It's not a fluke. It's how Firefly interprets material language.
Then I Found Something Better
After confirming the A/B gap, I tested a third option: "transitions into."
Extreme macro photograph of leaf surface where cell structure
transitions into printed circuit board traces, copper pathways
replacing leaf veins, tiny LED lights glowing at vein
intersections, soft backlighting revealing translucent leaf
membrane, dark background, 100mm macro lens, shallow depth
of field, professional scientific photography, hyper-realistic,
ultra-detailed
Where organic ends and technology begins. The veins carry nutrients on the left side and electricity on the right. The backlighting makes the translucent leaf membrane glow around copper traces. Firefly even added unprompted SMD chip components at junction points.
This scored 9.07 average - beating "constructed from" at 8.91.
The difference is subtle but important. "Constructed from" produces a replacement: the wing IS clockwork. "Transitions into" produces a gradient: the leaf is becoming a circuit board. You can see where organic ends and technological begins, but the boundary is fluid. Part of the leaf is still green, still photosynthesizing. Part is copper traces and glowing LEDs.
And the reason it works so well here is that leaf veins and circuit traces are functionally the same thing. Both are branching networks that carry resources from a central trunk to distributed endpoints. Veins carry nutrients. Traces carry electricity. Firefly saw that parallel and exploited it - the veins didn't just become copper, they became copper traces that follow the exact same branching logic.
The Rule I Didn't Know I Was Testing
The macro session accidentally revealed a deeper principle. Look at the scores:
Subject to Material | Structural Mapping | Score |
|---|---|---|
Leaf veins to copper circuit traces | Branching network to branching network | 9.07 |
Wing veins to brass rod framework | Support structure to support structure | 8.91 |
Pomegranate seeds to ruby jewel bearings | Round objects in chambers to round objects in settings | 8.88 |
Rose petal curves to crystal facets | Soft curves to hard angles | 7.60 |
Three of these worked. One didn't. The difference isn't the material - it's whether the subject has features that map onto features of the material.
I'm calling it the Structural Mapping Principle: material transformations succeed when you can name which part of the subject becomes which part of the material. Wing veins become brass rods. Seeds become jewel bearings. Leaf veins become copper traces. Each source feature has a destination feature.
Rose petals? They have curves. Crystal has facets. There's no mapping. So the crystal just... became crystal. The rose disappeared entirely. Four beautiful images of quartz crystal specimens. Zero images of crystal flowers.
Gorgeous crystal. Spectacular prismatic rainbow bands. But where's the rose? I asked for a rose petal constructed from crystal. Firefly gave me rutilated quartz. Beautiful, technically flawless, and zero percent flower. When there's no structural mapping, the material wins and the subject vanishes.
The Clockwork Pomegranate (And Why Interior Revelation Works)
The test I was most uncertain about was slicing open a pomegranate to reveal clockwork inside.
Extreme macro cross-section of pomegranate revealing intricate
clockwork mechanism interior, brass gears replacing seeds,
ruby-colored jewel bearings in each seed chamber, soft side
lighting, turquoise gradient background, 100mm macro lens,
focus stacking sharp throughout, professional product photography,
hyper-realistic, ultra-detailed
9.50 - the highest single image of the session. Rind becomes case. Chambers become compartments. Seeds become ruby jewel bearings in silver settings. Every feature of the fruit has a mechanical analog, and Firefly exploited every one of them.
The structural mapping is almost too perfect. A pomegranate's anatomy already looks like a watch interior: chambers housing round objects arranged around a central core. Firefly made the conceptual leap that pomegranate seeds are already ruby-red and gem-like, so it replaced them with actual watchmaker's jewel bearings in silver settings. The rind became the case. The central pith became the mainspring barrel. One image even had watch hands visible - suggesting this isn't just gears inside a fruit, it's a functional timepiece wearing a pomegranate as a disguise.
When Firefly Says No
Not everything worked. The water droplet test was supposed to show a miniature mountain landscape inside a dewdrop - a world within a world. Instead, Firefly gave me physically accurate refraction. The droplets acted as tiny lenses showing inverted views of the surrounding grass and sky, which is exactly what real water droplets do.
Technically flawless macro photography of real water droplet physics. The sunset catches the droplet edge perfectly. But I asked for a mountain landscape inside the drop, and Firefly gave me mostly refracted grass instead. Physics won.
I've now catalogued three ways Firefly rejects impossible prompts:
Training data shortcuts. When a concept maps to an existing Stock photography category, Firefly takes the shortcut. "Tiny elephant on saucer" becomes figurine product photography instead of impossible biology. (I documented this in Part 1.)
Material dominance. When the subject and material have no structural mapping, the material wins. "Crystal rose petal" becomes crystal. The rose vanishes.
Physics override. When the impossible element contradicts well-established physics in the training data, Firefly defaults to reality. Water droplets refract - Firefly knows this from thousands of macro photographs - and no amount of "containing a mountain landscape" overrides that knowledge.
Each failure mode has the same root cause: Firefly reaches for what it knows. When what it knows conflicts with what you're asking for, its knowledge wins.
What This Means for Your Prompts
Three actionable takeaways:
Use "constructed from" for complete material replacement, "transitions into" for transformation gradients. "Made of" is the worst option in every test I've run. Drop it from your vocabulary. If your subject and material share a functional structure (things that branch, things that contain, things that frame), "transitions into" will beat "constructed from." If you want full commitment to the new material, "constructed from" is your tool.
Check the structural map before you start. Ask yourself: "Can I name which feature of the subject becomes which feature of the material?" If you can list three or more feature-to-feature mappings, you're in good shape. If you can't name any, pick a different pairing. This single diagnostic question would have saved me from the crystal rose failure.
Respect Firefly's knowledge. The three failure modes aren't bugs - they're Firefly doing what it was trained to do. You can work with that knowledge (use "constructed from" to override training shortcuts) or you can run into it (ask water to not refract). The goal is finding the boundary between "impossible enough to be interesting" and "so impossible that Firefly falls back to reality."
The Numbers
Variation | Language | Avg Score | Best Score |
|---|---|---|---|
Leaf to Circuit Board | "Transitions into" | 9.07 | 9.35 |
Clockwork Butterfly | "Constructed from" | 8.91 | 9.35 |
Clockwork Pomegranate | "Constructed from" | 8.88 | 9.50 |
Clockwork Butterfly | "Made of" | 7.93 | 8.25 |
Crystal Rose | "Constructed from" | 7.60 | 7.90 |
Landscape in Droplet | Scene-in-scene | 7.40 | 7.80 |
24 images. 6 prompt variations. Three languages tested. One principle discovered.
Map the structure. Choose the language. Let Firefly engineer the rest.
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.