I had a mirror problem.

Not a personal one, though my wife might disagree. A prompting one. I'd been testing impossible reflections in @adobefirefly, asking mirrors to show worlds that couldn't possibly be there, and the results were consistently disappointing. Every mirror image came back looking like someone had taped a poster to the wall behind the subject. Flat. Lifeless. No reflection physics at all.

The bathroom mirror variation averaged 7.67. Not bad in isolation, but compared to the puddle variation at 8.09 and the glass window at 8.95, something was clearly wrong. Firefly understood "wrong reflection in a puddle" perfectly. It just couldn't figure out "wrong reflection in a mirror."

So I ran a controlled experiment. Same concept (impossible reflection), same mirror surface, same bathroom setting. The only difference was the language I used to describe the mirror itself.

Here's the original mirror prompt, stripped to the relevant section:

And here's what Firefly produced:

See that? The jungle is just sitting there. Behind the person. Like a backdrop at a department store portrait studio. There's no glass surface. No reflection sheen. No interaction between the mirror world and the real world. Firefly treated "mirror" as "window" and "reflection shows" as "background is."

Four out of four images did this. It wasn't a fluke.

I added three types of language to the next mirror prompt. Not new concepts. Just descriptions of things mirrors actually do in real life.

"Glass surface of mirror slightly fogged at edges."

That's it. That's the headline. But let me explain why those words matter so much.

Fog can't exist on a non-physical surface. By telling Firefly the mirror has condensation, I forced it to acknowledge the mirror as an actual object with material properties. Glass. Temperature. Moisture. Suddenly it's not a transparent window anymore. It's a surface that things can cling to.

The second addition: "Cool blue cosmic light from the reflection illuminating her face and white bathroom tiles."

This told Firefly that the impossible world inside the mirror casts light outward into reality. That's the bridge between the two worlds. Without it, the reflected scene just sits there, isolated. With it, the scene reaches out and touches the real environment. Light is the handshake between real and surreal.

The third: "Her face reflected on the glass surface with nebula visible behind her reflection."

This explicitly described a layered image. Glass surface on top. Face reflected on that surface. Nebula behind the face. Three depth layers. Without this, Firefly sees two layers: person, background. With it, it sees three: surface, reflection, other world.

Here's the full fixed prompt:

The unfixed mirror prompt averaged 7.67. The fixed version averaged 8.51. That's a 0.84 point improvement, roughly 12%, from language changes alone. Same mirror. Same bathroom. Same concept. Different words.

But the numbers only tell part of the story. Look at what actually changed:

The fog showed up every single time. Four out of four images rendered visible condensation on the mirror surface. Not once did the original prompt produce anything on the glass.

The light casting worked in every image too. Blue cosmic light from the nebula spilled onto the woman's face and the bathroom tiles around her. In one image, the peak scorer at 8.88, her face was semi-transparent, layered over the galaxy like a double exposure. You could see her reflection ON the glass and the nebula THROUGH it simultaneously.

That image doesn't exist in stock photography. I checked. An elderly woman's face ghosted over a spiral galaxy, rendered with the naturalism of a bathroom snapshot? That's new. And it happened because three lines of prompt text told Firefly to treat a mirror like a mirror.

Here's my theory: Firefly's training data includes millions of stock photos with mirrors in them. But almost none of those mirrors show impossible scenes. So when you ask for "mirror reflecting a jungle," Firefly doesn't have a reference for that specific combination. It falls back on the simplest visual approach. Composite the scene behind the person.

But fogged mirrors? Steamy bathroom mirrors with condensation? Those exist all over stock photography. "Person looking into fogged mirror" is a well-represented genre. By adding fog language, I didn't teach Firefly something new. I triggered a genre it already knew. And that genre happens to include all the surface physics, glass sheen, moisture, light interaction, that the basic "mirror reflection" prompt was missing.

The same logic applies to the light casting. Stock photos of people lit by screens, TVs, and phone screens at night are everywhere. "Face illuminated by blue light from glass surface" is a pattern Firefly has seen countless times. I just redirected that pattern into a surreal context.

This mirror fix taught me something bigger than mirrors. It's a principle that showed up across my entire impossible reflections session:

Describing surface imperfections forces AI to render physical surfaces.

Fog on mirrors. Fingerprints on glass windows. Tarnish and patina on polished metal. Every time I described how a surface was imperfect, aged, dirty, fogged, smudged, Firefly responded by rendering an actual physical surface instead of a flat composite. And every time it rendered a physical surface, the scores went up.

The same principle worked on a silver hand mirror. Adding "tarnish and patina visible on silver frame" produced museum-quality aging on the metal, green verdigris corrosion, centuries of wear in the filigree. And the reflected storm inside that tarnished mirror hit 9.38, the highest score of the entire session.

I think there's a reason this works so consistently. AI image generators don't understand physics. They understand visual patterns. A "fogged mirror" triggers a specific visual pattern that includes glass surface properties, condensation droplets, and light refraction through moisture. A "mirror reflecting X" triggers a different pattern that's essentially just a frame with a different image inside it.

You're not teaching the AI physics. You're choosing which visual pattern it follows. And patterns with physical imperfections tend to be richer, more detailed, and more photorealistic than clean, idealized ones.

Next time you're prompting any AI image generator and you want a reflective or glass surface to feel real, describe its imperfections. Don't just say "mirror." Say "slightly fogged mirror." Don't just say "glass window." Say "cold glass with breath fog and fingerprints." Don't just say "polished metal." Say "tarnished silver with patina."

Three words. "Fogged at edges." That was the difference between a 7.67 and an 8.51.

Sometimes the most powerful prompt engineering isn't about describing what you want to see. It's about describing what you want to touch.

Part 2 of the Impossible Reflections series. Part 1 covered the surface hierarchy, why puddles, mirrors, and glass all produce different types of impossible reflections. Coming next: why subtle contrasts fail in AI, and the minimum "impossibility threshold" your reflected scene needs to clear.