The Seven-Clause Formula for Getting Gemini to Shoot Neon Noir Like a Film Director.

Seven things happening in that prompt. Seven decisions. Each one is doing something specific — not "for vibes," but because it pulls the model's probability distribution toward a particular output. Here's what each clause is actually doing.

Featured Prompt

Clause Breakdown

"Atmospheric neon noir street scene"

This is a genre declaration, and genre declarations are the most efficient tokens in any prompt. "Neon noir" activates a specific cluster of visual training data — Blade Runner, the entire catalog of late-'80s Japanese cinema, a thousand moody urban photography portfolios. That cluster has strong consensus on color temperature (cool shadows, warm-to-saturated neons), time of day (always night), and weather (almost always rain or fog). You get a lot of data for two words.

"Rain-soaked Tokyo alley at 3am"

Tokyo is working harder than "city" here. Gemini's training data has categorical differences between "Tokyo alley," "Hong Kong alley," and "Manhattan alley." Tokyo pulls narrow pedestrian lanes, lantern-style signage, dense vertical surfaces, temple gate aesthetic bleeding into urban context. The "at 3am" time specification biases toward depopulated scenes — which is what you want if you're putting a lone figure in frame. Put "at 9pm" and you get crowds competing with your subject.

"Neon izakaya signs reflected in wet concrete pavement"

The reflection demand is what sells rain. "Wet pavement" without "reflected" produces surface sheen at best. "Reflected in wet concrete pavement" forces the model to render the actual light source twice — once above, once below — which produces the depth and glow that defines the look. "Izakaya" is more specific than "restaurant signs" and pulls toward the kanji-heavy, rectangular format of Japanese bar signage, which reads differently than Times Square or Kowloon in the output.

"Atmospheric perspective with shallow foreground depth"

Two depth instructions in one clause. "Atmospheric perspective" asks for aerial haze — objects becoming less defined, lower contrast, slightly warmer-tinted as they recede. "Shallow foreground depth" simulates a wide-aperture lens — a slight blur on elements closest to camera. Together they create a three-plane composition (crisp mid-ground subject, soft foreground, hazy background) that reads as cinematic without ever mentioning "cinematic." The c-word is overloaded in AI prompts and now produces inconsistent results across Gemini versions.

"Grain and halation consistent with 35mm Fuji 400H pushed to 1600"

This is the most technically specific clause and the one with the highest variance. "Film grain" gives you digital-looking grain. Naming a specific stock (Fuji 400H) pulls from a well-documented photography subculture that catalogued exactly what that film looks like — slightly green shadows, lifted blacks, warm-to-neutral rendering. The push-process notation tells the model you want denser grain and compressed dynamic range. Not every Gemini run lands accurate Fuji 400H — but naming it shifts the output away from generic AI smoothness toward something that reads as photochemical.

5 Variations to Test

Hong Kong density: Swap Tokyo alley for "Hong Kong Mong Kok side street at 3am" and add "layered signage extending three stories above street level." Hong Kong training data skews toward much more vertical, chaotic neon stacking. You lose the Japanese minimalism and gain a more overwhelming composition.

Monochrome with single neon accent: Add "monochrome base palette with single point of color: electric teal neon tube" after the color description. Gemini desaturates the ambient and isolates the neon. Different emotional register — more neo-noir than neon noir.

Interior version: Replace the street with "rain visible through steamed plate-glass window of jazz bar, viewed from exterior sidewalk, upright bass player in soft spotlight inside, condensation distorting the glass." The reflection logic shifts to the glass surface. More intimate output, works better for editorial use.

Push the grain harder: Swap Fuji 400H pushed to 1600 for "Kodak Tri-X 400 pushed to 3200" for a grittier, more contrasty output. Tri-X pushed 3 stops produces visible grain clusters rather than smooth film grain.

Daytime color study: Replace "3am" with "overcast afternoon" and remove neon references. You get a greyer, quieter scene — but the atmospheric perspective and depth clauses still work, testing which elements are doing lighting work vs. compositional work.

Model Compatibility

Gemini Imagen 3 (Nano Banana Pro): Excellent. Best handling of complex lighting and reflection demands. Film stock reference lands on ~70% of runs.

Midjourney v6: Excellent. Add `--style raw` to reduce oversaturation; the model defaults to punchy neon that overshoots the muted base palette. Lower `--s 150` preserves charcoal tones.

DALL-E 3: Good. The film stock reference rarely produces accurate halation. Replace with "analog film grain with slight light bloom on brightest elements." Remove "anamorphic flare" — DALL-E renders it inconsistently.

Stable Diffusion XL: Good. Needs Film Grain LoRA for the texture. Depth instructions work well with DPM++ 2M Karras at 25+ steps. Add negative: "neon bloom overexposed, oversaturated."

Failure Modes

When this prompt fails, it usually falls into one of three patterns:

The generic cyberpunk — neon goes hot pink and electric blue, architecture becomes generic sci-fi, the figure disappears. This happens when the model over-weights "neon noir" at the expense of the depth instructions. Fix: move atmospheric perspective earlier in the prompt.

The flat pavement — rain is implied by sheen but there are no actual reflections. Fix: add "mirror-like reflection of every light source in the wet pavement" as a separate clause.

The overlit figure — the raincoat figure becomes the brightest element instead of a mid-tone. Fix: add "figure in semi-silhouette, ambient-lit only, no direct light on figure."

Share Your Runs

Run this. Change one clause. See what breaks and what improves. That's the whole game with AI prompting — it's empirical, not intuitive.

Drop your results in the comments or tag @chickenpie.co. Especially interested in seeing what Midjourney does with the film stock reference vs. Gemini.