Most of my research tracks don't have a destination. I run controlled sessions, score the outputs, document the findings, and publish what I learned. The work is the point. Temporal portals were different.

I built this concept specifically for Adobe's "How I Firefly" competition in late 2025. That changed everything about how I made decisions. Not the methodology. The methodology stays the same regardless. But the filter I ran every choice through shifted from "what's most interesting?" to "what will hold up in front of a judge?"

That's a different question. And it produces different answers.

46 images. 3 sessions. 3 hours. Here's how competition thinking shaped the process from concept to submission.

The first competition decision happens before you generate a single image: what are you actually trying to make?

"Weird but photographic" is the aesthetic I always chase, and it's a useful filter for competition work because it describes something specific. Not abstract. Not illustrative. Not concept art. Photographic, meaning the impossible thing is rendered with the technical precision of real optics, real light, real physics.

The gravity well portrait checked that box clearly. A person standing in a contemporary room while visible gravitational distortion pulls objects toward a circular portal, and through that portal, a different time period. Present and past in a single frame, connected by physics you could see.

The scenario is impossible. But gravitational lensing, atmospheric depth, and light refraction are all documentable phenomena in real photography. The concept gave Firefly something to be right or wrong about, technically. That matters in competition. Judges can recognize when physics is rendered convincingly. They can also recognize when it isn't.

My target score going in: 90 or above on a 100-point scale. I'd never hit that. The highest single image from any previous research track was 9.85 on the 10-point scale, roughly equivalent, but that was creature photography with a clear subject. This was conceptual, atmospheric, physics-dependent. Higher degree of difficulty. The right kind of difficult for a competition entry.

Six variations, 24 images, one goal: understand which elements the concept actually needed before refining anything.

I tested across time periods first. A 1950s kitchen portal gave me an 8.48 average, competent execution of the concept, the fabric-flowing-toward-portal gravity effect working well, but nothing that stopped you. Art Deco 1920s scored similarly at 8.45 average; the era's visual language is apparently harder for Firefly to render distinctly, tending to blur into generic vintage.

Victorian and Medieval were different. Two variations, a Victorian study with spiral objects and a medieval great hall with dramatic architecture, both peaked at 9.85. Not average: individual images. The Victorian spiral prompt (Variation B) produced Image B-1, which I scored 10/10 on Visual Quality, Prompt Alignment, Uniqueness, and X Engagement. It was the first image of the session that looked like it could actually win something.

What these two eras share: architectural density. Floor-to-ceiling bookshelves. Stone vaulted ceilings. Gas lamps and chandeliers. Gothic arches. When the portal destination has that much visual complexity and clear historical identity, the temporal contrast becomes immediately readable. You don't have to tell the viewer they're looking at two different times. The image tells them.

Session 1 average: 8.69/10. Strong baseline. Not competition-ready.

The pattern that emerged: spiral gravity visualization, not random floating objects, but objects following a precise curved mathematical trajectory, was the single most important structural element. Variations without it scored 5 to 7 points lower. Gravity needs to look like gravity, not like someone hit zero-G and things drifted. In a competition context, that distinction is everything: vague physics reads as decoration, precise physics reads as craft.

With Victorian and Medieval identified as the strongest time periods, Session 2 was refinement. Four variations, 16 images, all targeting 90+.

The first attempt, adding specific named objects (leather-bound books, brass compass, pocket watch, fountain pen), Rembrandt lighting, and "museum-quality composition" language, actually regressed. Average dropped to 76.93, best image at 78.30. Lower than my session 1 baseline.

This is the kind of result that only makes sense when you document failures carefully. "More specific and more refined" doesn't always produce better images. The Rembrandt lighting direction competed with the portal glow direction. The specific object list created crowding without trajectory. The "museum-quality" language pulled Firefly toward fine art stillness rather than dynamic physics visualization. Adding complexity in the wrong places degraded the core concept.

The second Victorian attempt dropped Rembrandt and added two things I hadn't tried: atmospheric haze ("atmospheric haze revealing light rays") and prismatic light refraction along the spiral trajectory. The score jumped to 88.90 on the best image, an 11-point improvement over the regression, a 10-point improvement over session 1 baseline.

That's when I knew I'd found the technique.

Atmospheric haze is something photographers understand and AI generally doesn't do well. Real long-exposure photography, real interior shots with window light: the air in those images has visible density. Dust. Humidity. The light doesn't just travel to the lens; it scatters, and that scattering creates depth layers that pure digital rendering eliminates. Asking for atmospheric haze was asking Firefly to do something that required understanding what makes space feel inhabited rather than rendered. When it worked, the portal didn't float in a clean digital void. It glowed through air.

The medieval version pushed the score another notch to 89.85, using fire and smoke for atmospheric depth instead of haze. Volumetric firelight through the spiral. Jaw-dropping, technically.

And here is where competition thinking diverged from pure research thinking.

On raw score, medieval won. 89.85 vs. 88.90. If this had been a pure research track, I would have taken medieval further and documented why fire and smoke outperformed atmospheric haze.

But this wasn't pure research. Adobe's "How I Firefly" competition is judged by people who think in photography. The medieval image looked like VFX. Spectacular, technically impressive, immediately recognizable as AI doing something dramatic. The Victorian haze image looked like a composite photographer who understood the physics had spent weeks on a single frame. Sophisticated. Harder to dismiss.

In competition, the image that's harder to dismiss is the image that wins.

I went with Victorian.

One prompt. Six generations. Find the submission.

Results across 6 generations: 92.60, 84.85, 88.00, 92.10, 83.50, 91.45.

The variance in that list is the real lesson. The same prompt produced a 92.60 and an 83.50 in the same session. That 9-point range is why the methodology calls for minimum 4 generations per variation: any single image from this session would have been misleading data. The 83.50 would have underrepresented the formula. The 92.60 in isolation would have looked like a fluke.

In competition, you can't submit a fluke. You have to run enough generations to know which image is representative of the formula and which one is an outlier in either direction.

Generation 1 scored 92.60. Competition submission. What distinguished it:

The prismatic light refraction was calibrated exactly right, concentrated along the spiral trajectory, not diffused across the entire frame. When prismatic effects dominate everything, they become abstract. When they're targeted to the physics demonstration, they become evidence. This image had both: the atmospheric haze creating depth in the room, the prismatic refraction proving the spiral's physics, the Victorian library sharp and detailed beyond the portal, the contemporary loft window establishing the present. Every element doing specific work.

It didn't look like AI art. It looked like a composite photographer who understood gravitational lensing had spent weeks on a single image. That's the competition standard I was building toward from session one.

The score progression across three sessions:

Session 1 average: 8.69/10 → Session 2 breakthrough: 88.90/100 → Final winner: 92.60/100

Breaking that down by technique contribution:

Mathematical spiral pattern (not random floating): approximately +25% improvement over generic gravity distortion. Objects need to follow a curved trajectory that demonstrates physics, not just drift.

Atmospheric haze: +6 to 8 points. The biggest single technique discovery in this research track. "Atmospheric haze revealing light rays" is four words that tell Firefly to make the air visible, and visible air is what separates photographic depth from digital sterility.

Prismatic light refraction along trajectory: +3.7 points. Targeted application matters. "Creating prismatic effects" everywhere weakens the image. "Light refraction visible along spiral trajectory creating prismatic effects" tells the model where the physics is happening.

Historical era specificity: "Victorian library with floor-to-ceiling bookshelves and ornate writing desk under chandelier" vs. "old room." Named architectural elements render with conviction. Vague period language blurs into generic vintage.

Deep focus over shallow: Both time periods sharp. Counterintuitive for portrait photography, essential for temporal storytelling. Shallow DOF loses the historical destination. The contrast between present and past requires both to be visible.

Historical era rankings from testing:

Victorian (library/study): Most consistent. Architectural density renders beautifully: floor-to-ceiling bookshelves, ornate writing desk, chandelier, gas lamps.

Medieval (castle/great hall): Maximum drama. Stone archways, vaulted ceiling, burning braziers, crimson banners.

1960s (living room): Good period details, less architectural grandeur.

1950s (kitchen): Recognizable, scores lower on uniqueness.

Art Deco 1920s: Avoid. Firefly renders this indistinctly.

Success rate with refined formula: 86% of variations score 80+. Three of the six final-session generations scored 90 or above.

The atmospheric haze discovery is the technique I keep returning to across other research tracks. But the bigger lesson from this project isn't about a prompt technique.

It's about the filter.

When you're building for competition, every decision has a second question attached to it. Not just "does this score higher?" but "does this hold up in front of someone who knows photography?" Those aren't always the same question, and the gap between them is where competition entries get made or lost.

The medieval image scored higher. I submitted Victorian. That choice only makes sense if you're thinking about who's judging and what they're actually looking for. Spectacle is easy to recognize and easy to dismiss. Craft is harder to explain, which means it's harder to argue with.

For the technical side: "atmospheric haze revealing light rays" is four words that tell Firefly to make the air visible. It costs you nothing. But it signals to the model that you understand what makes space feel real rather than rendered, and that signal changes the register the image is executed in.

I've since tested this technique across multiple research tracks. It doesn't work everywhere, but wherever the concept involves dramatic lighting through interior space, it's worth trying. The points it adds are the difference between an image that looks good and one that looks like it was made.

Start here, swap the historical era, add architectural specificity:

Name your objects. Name your architecture. If you're using Victorian, say floor-to-ceiling bookshelves, not books. Say ornate writing desk under chandelier, not furniture. Specificity is physics. The more precisely you describe the destination, the more convincingly Firefly renders the impossible journey there.

Testing conducted in Adobe Firefly Image 5. 46 images across 16 variations and 3 sessions. Scored using a weighted five-dimension rubric: Visual Quality (30%), Prompt Alignment (25%), Consistency (15%), Uniqueness (15%), X Engagement Potential (15%). Competition entry scored 92.60.