Part 1 established the rules. Language precision. Material physics. Gradient backgrounds. Rescue strategies. Fire + Ice as the universal champion at 9.85. Coral breaking the organic curse at 9.20. A material tier list built from 128 scored images.

But every one of those images was a portrait. Same framing. Same subject type. Same face staring back at me from a gradient background.

The question Part 2 answers: do the patterns hold when the subject changes?

Hands. Full bodies in motion. Classical sculpture. Real-world environments.

The answer is yes. But the surprises along the way rewrote my understanding of what makes AI material transformation actually work.

Hands are the anatomy lottery. Everyone who generates AI images knows this. Fingers merge. Extra thumbs appear. Knuckles bend wrong. Material transformation on hands should make the problem worse because now the AI is rendering unfamiliar materials AND unfamiliar anatomy simultaneously.

I tested Fire + Ice on hands playing a piano. Left hand blazing with flames, right hand crystalline ice, keys reacting to both.

Peak score: 8.85. Set average: 8.49.

Lower than portraits (9.43). But here's the finding that matters more than the score.

The piano keys fixed the fingers.

Props constrain finger placement. When five fingers need to press five keys, the AI has structural guidance for where each digit goes. No extra thumbs. No merged fingers. The object tells the AI what hands are supposed to do.

And the keys did something else. Fire-side keys warped slightly from heat. Ice-side keys developed frost crystals on the ivory. The object reacted to the materials. A charred key and a frosted key prove the transformation is physically real in ways that a gradient background never can.

The slight score dip versus portraits came from reduced visual drama at macro scale. Hands are smaller canvas than faces. But the trade was worth it: narrative depth that portraits can't deliver.

Glass + Bronze scored 9.21 on portraits. Clean left/right split. Glass with prismatic refraction on one side, polished bronze with green patina on the other. The transition zone where transparency becomes opacity was the visual payoff.

I put that combination on a dancer mid-leap.

The split never happened. Not in any of the four images. Firefly could not maintain a clean 50/50 material boundary on a body in motion. Crossing limbs, foreshortening, asymmetric poses. Too many variables. The AI defaulted to easier solutions: a bronze body with a glass tutu. A glass lattice overlay on bronze. A bronze skeleton visible through glass skin.

That last one scored 8.80.

The glass-skeletal overlay was an accidental discovery. Firefly found that layering glass transparency over bronze to reveal internal anatomy is more interesting than what I actually prompted. You could see bronze bone structure through transparent glass muscle. An X-ray rendered in precious materials.

The score gap between the dancer (8.80) and the portrait (9.21) came entirely from reduced prompt alignment. Visual quality and engagement potential actually scored higher than portraits. Dance photography is inherently more dynamic. The AI produced magazine-cover compositions that a static portrait never could.

But it solved the wrong problem. Or maybe the right one.

Material splits work on static subjects. Dynamic poses break the boundary. And sometimes what the AI gives you instead is better than what you asked for.

Here's something I didn't fully understand until 128 images in.

Every face I'd generated had the same tension: the materials were fighting the face. Glass wanted to be rigid. Fire wanted to consume. Coral wanted to grow. But the face needed to stay human. Recognizable expressions. Readable emotions. The prompt was always asking for two contradictory things: be this material AND be this person.

Some materials handled the tension well. Fire + Ice acting ON a face preserved the person underneath. The materials were a transformation in progress. Something happening to a human being.

But replacement materials like glass, coral, and obsidian turned the face into a mannequin. The face IS the material. No person underneath. Beautiful, but lifeless. A sculpture, not a portrait.

That distinction between transformation and replacement had been quietly shaping every score in the project. And the solution came from leaning into the problem instead of fighting it.

What if the subject is supposed to be a sculpture?

Classical Greek sculpture bust. White marble with visible veining. Laurel wreath crown. Museum photography lighting. Dark charcoal gradient background.

And the marble is actively transforming into liquid mercury. Top of the head still solid stone. Face dissolving into mirror-polished liquid metal. Mercury pooling and dripping from the chin with visible surface tension. The laurel wreath half stone, half flowing silver.

Set average: 9.44. The highest in project history.

Peak score: 9.85. Tied with Fire + Ice for the all-time record.

But here's what made this different. The weakest image in the set scored 8.85. In every other variation, the weakest image dragged the average down by a point or more. Mercury bust's floor was higher than most variations' ceilings.

Three things converged to produce this result.

Art historical context is a score multiplier. Marble busts are among the most recognizable art objects in visual culture. Firefly's training data is dense with classical sculpture references. The AI knew exactly what a Greek bust looks like, which meant every generation started from a stronger foundation than an arbitrary human face.

Sculpture subjects eliminate the mannequin problem. When the subject is supposed to be a sculpture, looking sculptural becomes a feature instead of a bug. No need to maintain a living expression. The material transformation IS the life in the image. A monument dissolving. A permanent thing becoming impermanent.

Mercury's liquid physics create implied time. Unlike frozen material splits, mercury drips, pools, flows. Every image captures a process happening right now. Viewers extrapolate backward (the bust was intact) and forward (it will be consumed). That implied timeline turns a single frame into a three-act story: what was, what is, what will be.

The emotional resonance surprised me most. I-4 showed mercury "weeping" down the face from the crown like a metallic veil. A monument mourning its own dissolution. I didn't prompt for emotion. The materials created it.

The mercury bust on a gradient scored 9.85 peak and 9.44 average. Exceptional. But it was a blob of liquid metal dissolving a marble head on a featureless background. A technical demonstration.

So I placed it in a dimly lit museum gallery at night. Single dramatic spotlight from above. Other sculptures barely visible as dark silhouettes in the background. Polished gallery floor reflecting the mercury and marble. A contemporary art installation dissolving after hours.

Same subject. Same materials. Different context.

Peak score: 9.85. Same as the gradient. But the set average climbed to 9.63, the highest in the entire project across all 164 images. And the range compressed from 1.00 to 0.40. The worst museum image (9.45) beat the worst gradient image (8.85) by 0.60 points.

The gradient version was a demo. The museum version was a story.

One detail I didn't anticipate: the warm museum spotlight turned the mercury gold. I prompted silver mercury. The AI rendered liquid gold. Because mercury is a mirror, and mirrors reflect their light source. Warm spotlight means warm reflections. The training data association of classical bust + museum + liquid metal pulled everything toward gilded sculpture.

That's not a bug. That's physics. Your lighting color temperature will override your material color specification for any reflective material. If you want silver mercury, give it cool light. If you want gold, give it warm light. The material reflects whatever you point at it.

[IMAGE: Mercury Classical Bust in museum, K-R1-4, scored 9.85 with 9.63 set average. The gradient version's twin with a higher floor.]

The museum bust outperformed the gradient bust. That contradicted everything I'd established in Part 1. Gradients win by nearly a full point over environments. The winter cafe proved it. Fire + Ice dropped from 9.20 on a gradient to 8.25 in a real-world setting.

But the mercury museum matched and exceeded the gradient.

Two environmental tests. Opposite results. One material loses a point in context. Another gains consistency.

What was different?

The answer turned out to be the most important discovery in the entire project. It wasn't about environments versus gradients. It was about something far more specific. Something I couldn't see until I spent an entire session testing it.

Part 3 is about that discovery. It changes how you think about every material transformation prompt you'll ever write.

The question is simple: who controls the light?

Four new subject types. 40 additional images. And the single highest-performing set in 164 generations.

Part 3 figures out what.

This is Part 2 of "164 Images Later," a 3-part series on systematic AI material transformation testing. Part 3 drops next week.

Every prompt shared in full. Every score documented. The prompt companion with generation notes is linked below.

Built with Adobe Firefly Image 5.