I went into this thinking I was testing one concept. A puddle that reflects a world that isn't there. A mirror showing something impossible. Simple, right? One idea, different surfaces.
Twenty-four images later, I'd accidentally discovered that @adobefirefly doesn't have one method for handling impossible reflections. It has three. And which one it picks depends entirely on the surface you give it.
The Test
I generated four images each across six variations: rain puddles, bathroom mirrors (two versions), a store window, and a polished silver hand mirror. Every prompt asked for the same basic thing. A reflective surface showing a scene that couldn't possibly be there. Greek temples in city puddles. Tropical jungles in bathroom mirrors. Underwater oceans in shop windows. Storms trapped inside hand mirrors.
Same concept. Same model. Same intent. Wildly different results.
Same concept. Four surfaces. Four completely different AI interpretations.
Trick #1: The Inverted Reflection (Puddles)
When I put a Greek temple in a rain puddle, Firefly did exactly what I expected. It rendered a proper upside-down reflection. The temple inverted in the water, golden light bouncing off wet pavement, the whole thing obeying actual reflection physics. The water surface behaved like water. The image scored 8.46 at its peak.
Here's the prompt:
Businessman in modern suit standing on wet city sidewalk at night,
rain puddle at his feet reflecting a completely different scene
showing grand ancient Greek temple with marble columns and warm
golden lantern glow, reflection is sharp and detailed, street
lights creating wet pavement glow, 35mm street photography,
shallow depth of field focused on puddle reflection, hyper-realistic,
cinematic lightingThis makes sense when you think about it. Adobe Stock is full of creative puddle reflection photography. Photographers love shooting cities reflected in rain puddles, autumn leaves mirrored in still water. Firefly has seen thousands of these. It knows what a "wrong" puddle reflection looks like because the format already exists in its training data.
Peak score: 8.46. Firefly rendered true inverted reflection physics. Adobe Stock taught it how.
Trick #2: The Flat Backdrop (Mirrors, Before the Fix)
Mirrors were a different story. When I asked for a tropical jungle in a bathroom mirror, Firefly gave me wallpaper. The jungle sat flat behind the subject like a green screen. No reflection physics. No mirror surface sheen. No light bouncing between worlds. Just a guy standing in front of what looked like a themed shower curtain.
The prompt:
Man in white t-shirt looking into bathroom mirror, mirror
reflection shows dense tropical jungle with vines and exotic
birds instead of bathroom behind him, his face visible in both
reality and reflection, warm bathroom lighting contrasting cool
green jungle light from mirror, 50mm portrait lens, deep focus
showing both face and mirror reflection sharply, professional
photography, hyper-realistic detailAnd here's what's interesting. It was consistent. Four out of four images did the exact same thing. Firefly wasn't confused. It had a clear interpretation of "jungle in mirror." That interpretation just happened to be wrong.
The average score dropped to 7.67. Not terrible, but when you compare it to the puddle's 8.09 average, the gap tells a story. Mirrors, at least without intervention, produce the weakest version of this concept.
Average score: 7.67. Four out of four: wallpaper, not reflection. Consistent failure is still data.
Why the Difference?
Stock photography, probably. "Creative puddle reflections" is an entire subgenre on Adobe Stock. Photographers have been shooting those for decades. But "bathroom mirror showing a completely different world"? That's not really a thing in stock photography. Firefly doesn't have a strong reference point for it, so it defaults to the simplest compositing approach: paste the scene behind the person.
Trick #3: The Transparent Portal (Glass)
This one surprised me the most. I asked for an underwater ocean reflected in a store window, and Firefly didn't create a reflection at all. It created an aquarium.
Every single image shows a child looking through glass into an underwater world. Whale, coral reef, the works. The glass isn't reflective; it's transparent. The child isn't seeing a reflection of something impossible. She's looking through a window into another dimension.
And it scored 8.95 average. The highest variation average of the session.
The prompt:
Child pressing hands against shop window at night, low angle
shot from near ground level, fingerprints and breath fog visible
on cold glass surface, window reflecting a vast underwater ocean
scene with whale and coral reef instead of the street behind her,
blue-green underwater light from the reflection illuminating the
child's face and hands, neon store signs creating colored light
on glass surface, 35mm lens, deep focus showing both child and
reflection sharply, professional photography, hyper-realistic,
magical realism
Peak score: 9.08. The whale's nose meeting her hand through the glass. Firefly turned a reflection prompt into an aquarium.
And Then There's Metal
The silver hand mirror variation landed somewhere between puddle and portal. Firefly treated the polished metal surface as a contained frame, like a painting or a window into another place. The stormy ocean sits inside the mirror's oval frame, lightning crackling against waves, while the ornate silver filigree holds it all together on dark velvet.
It scored 8.98 average, with one image hitting 9.38. That was the highest single score of the entire session.
The prompt:
Close-up of ornate silver hand mirror lying on dark velvet surface,
tarnish and patina visible on silver frame, polished mirror surface
reflecting a stormy ocean with lightning and crashing waves instead
of the ceiling above, blue-white lightning flash from reflection
casting sharp light across velvet surface and silver frame, intricate
mirror frame details sharp and detailed, dramatic chiaroscuro lighting,
100mm macro lens, deep focus showing both mirror details and reflection
scene, professional product photography, hyper-realisticWhat made it work? Two things. First, removing the human subject entirely let Firefly focus on the object and its impossible contents. Second, the product photography framing (macro lens, dark velvet, dramatic lighting) gave Firefly a format it excels at. It's essentially still-life photography with a surreal twist. That's a format with clear training data precedent.
Score: 9.38. Session champion. A storm inside a mirror that lights up the table it's lying on.
The Surface Hierarchy
After scoring all 24 images, here's how the surfaces rank:
Polished metal objects scored highest at 8.98 average. Glass windows came in at 8.95. Mirrors with the right language (more on that in the next article) hit 8.51. Standard mirrors without adjustments sat at 7.67. And puddles, the format that actually produces true reflections, came in at 8.09.
The irony? The surface that creates the most technically accurate "impossible reflection" doesn't produce the best images. Puddles follow reflection physics correctly, but glass portals and metal-framed storms are more visually striking. Sometimes the wrong interpretation scores higher than the right one.
[IMAGE: Surface hierarchy visual - five surfaces with floating scores] The surface that renders "correct" reflections isn't the one that scores highest. Training data beats physics.
The surface that renders "correct" reflections isn't the one that scores highest. Training data beats physics.
What This Means for Your Prompts
If you're chasing this kind of impossible-reflection imagery, pick your surface based on what you actually want:
Want a true impossible reflection with proper inverted physics? Use a puddle. Want a portal into another world with maximum emotional impact? Use glass. Want a dark, atmospheric art object with the impossible contained inside a frame? Use polished metal. And if you want a mirror? You'll need the language fix I discovered, which completely changed how Firefly interprets the prompt. But that's a whole separate article.
[IMAGE: Final comparison strip - best from each surface: VA-4 puddle, VE-1 glass, VF-3 metal] Puddle reflects. Glass portals. Metal contains. Same prompt concept, three different AI behaviors.
Puddle reflects. Glass portals. Metal contains. Same prompt concept, three different AI behaviors.
The biggest takeaway from this session isn't about reflections at all, though. It's about training data. Firefly doesn't generate images from abstract understanding. It maps your prompt to the nearest format it's seen before. When you write a prompt, you're not just describing what you want. You're triggering a specific genre in Firefly's memory. Understanding which genre your prompt triggers is the difference between a 6.54 and a 9.38.
This article is part of an ongoing series where I systematically test @adobefirefly prompting techniques with scored results. Next up: the three-word mirror fix that improved scores by 12%.