Model Guide

How to Use GPT Image 2 for Cyberpunk World Key Visuals

A three-lesson workflow for turning world cards into key visuals: the city from a roof, a replicant portrait, and the rooftop ending.

How to Use GPT Image 2 for Cyberpunk World Key Visuals

This guide teaches the model through a world. Instead of a list of magic phrases, you get three lessons — a city, a character and a rooftop — each built from a card, a beat and a camera. By the end you should be able to write a cyberpunk key visual prompt in about four minutes and repair it in one pass.

It assumes no prior experience with image models and no encyclopaedic knowledge of Blade Runner. You need the world’s cards, GPT Image 2, and one decision per lesson: what is happening in the next ten seconds.

GPT Image 2

A blade runner in a long trench coat stands on a rain-slick rooftop above a neon Los Angeles, twin golden corporate ziggurats rising through brown smog behind him

GPT Image 2, world key visual: three readable planes — wet rooftop, crowded street canyon, corporate skyline — with a briefcase, spinners and a blimp holding the scale.

Krea 2

A lone coated figure with a briefcase looks out over a blue-black neon city with searchlights, a blimp and two gold towers

Krea 2 on the identical brief: the same city compressed into graphic bands, cleaner silhouettes, stronger thumbnail read and a much colder palette.

GPT Image 2

A dying platinum-haired replicant kneels in warm rain holding a white dove while a battered blade runner watches, a pistol abandoned in a puddle

GPT Image 2, event key visual: the dove, the nail, the scarred face and the abandoned pistol all survive in one legible frame.

Krea 2

A silver-haired man kneels on a rainy rooftop holding a white dove up to the rain while a dark-coated man stands opposite him

Krea 2 on the same rooftop beat: a flatter two-shot, the gun placed between the men like a prop in a stage play.

What You Need Before You Start

The world cards

Blade Runner: The Empathy Test has 8 characters, 6 locations, 6 events, 4 factions and 3 rules. You do not need all of them. You need one card, plus one rule that constrains it.

GPT Image 2 access

Text input is enough. Reference images are supported but not required for any lesson here.

Two frame sizes

1536×1024 for places and two-shots, 1024×1536 for single figures. Decide before you write, because the crop is part of the composition.

One exclusions line

“No text, no logos, no watermark.” Paste it every time. It prevents most of the boring failures.

Understand the World’s Visual Grammar

The three rules are also three image instructions. Learn them once and you stop describing the setting from scratch.

The Voight-Kampff Test → a machine between two people

A face held too still, a case on the table, a question nobody wants answered. If the frame includes the test, the machine is a compositional element, not a prop.

The Four-Year Lifespan → a body already running out

Still beautiful, already tired. A hand that will not stop shaking, a coat that is too heavy for the weather. The countdown should be visible in posture, not in a caption.

Memory Implantation → a domestic object that belongs to no one

A photograph, a piano, a childhood room. The object is ordinary; the ownership is impossible. Use it whenever a scene needs to feel intimate and false at the same time.

The grammar also has a colour rule: sodium and neon for the human world, one cold cyan or blue emission for the machine. Wet asphalt, warm amber practicals, magenta signage — then a cyan that does not come from the sky. If your image has no artificial cold light, it is a crime scene, not a cyberpunk scene.

Build the Prompt in Five Layers

Five layers, each doing a different job. If you only remember one thing from this guide, remember the order.

Layer 1: style and medium

High-contrast neon digital painting, hard-surface architecture, wet dark planes. This comes first or the detail clauses will override it.

Layer 2: subject and costume

Who, wearing what, holding what. Use the card’s exact nouns: trench coat, briefcase, Voight-Kampff case, platinum hair, black coat.

Layer 3: the three depth planes

Foreground, middle ground, background. This is the layer that separates a place from a wall of lights.

Layer 4: place, weather, named light

One place, one time of day, two named sources. Add rain as weather, not as a texture overlay.

Layer 5: camera and exclusions

Low or high camera, vertical or wide, no text, no logos, no watermark. This layer fixes most bad crops.

[style] Cyberpunk [shot type] key visual, high-contrast neon digital painting, [aspect]: [subject + costume + prop]; [foreground]; [middle ground]; [background]; [weather]; [key light + fill light]; [palette anchors]; [camera height]; no text, no logos, no watermark.

Lesson 1: Los Angeles from the Roof

Read the card

Los Angeles, November 2019: a rain-drowned vertical city of neon, steam, noodle stands and blimp light, ziggurat towers above permanent smog.

Choose the argument

One person is small inside a city that was not built for people. This is a scale decision, not a mood decision.

Write five layers

Neon digital painting; a blade runner with a briefcase, back to camera; wet roof and puddles; street canyon with stalls and spinners; ziggurats, smog and blimp; warm rain; magenta key plus amber fill; low camera, wide.

Generate at 1536×1024

Check the silhouette first, then the three planes, then the props. A beautiful image with a wall of lights is a failed establishing shot.

Repair one clause

If the city is flat, name the foreground object. If the figure is lost, raise the contrast of the coat against the street. Do not rewrite the whole prompt.

The GPT Image 2 result is a three-plane scene: rooftop, canyon, skyline. The Krea 2 result on the same prompt is a flatter poster with a bright blimp and two gold verticals. Both are valid. The GPT version explains the place; the Krea version sells it at a glance.

Lesson 2: A Replicant Portrait

Read the card

Roy Batty: platinum-haired Nexus-6 combat model, black trench coat, pale scarred face, animal eyes, a nail through one hand. The nail is the story; do not lose it.

Choose the beat

Not a fight. The moment after one, when he is tired and still standing. A character portrait is a pressure state, not a pose.

Write five layers

Neon digital painting; Roy in a black coat, nail through the left hand; rain, steam, a neon sign behind; cold cyan rim and warm amber fill; low camera, portrait.

Generate at 1024×1536

Check the hand first, then the face, then the coat. Hands are where the model hides its failures.

Repair the tells

If the scar disappears, name its position. If the nail becomes jewellery, describe it as “a nail driven through the palm, blood at the entry”.

The trap with this prompt is beautification. GPT Image 2 will make a dying man look like a cover model. Add “exhausted, not serene; rain on the face; no glamour lighting” and check at full resolution. The card’s word is “animal eyes” — if the eyes are calm, the portrait is wrong.

Lesson 3: Tears in the Rain

Reduce the beat to four facts

A dying replicant kneels. He holds a white dove. One hand is pierced by a nail. A blade runner watches, disarmed.

Stage the room

A rain-soaked rooftop ledge. Put the pistol in a puddle between them. The gun is the truce.

Write five layers

Neon digital painting; two men, one kneeling with a dove; wet ledge, vent pipes, neon below; warm night rain; magenta key and cyan rim; low camera, wide two-shot.

Generate at 1536×1024

Check the dove, the nail and the gun in that order. If any of the three is missing, the scene is a generic rainy argument.

Compare with Krea 2

The Krea version will be more symmetrical and more graphic. Keep it as the poster; keep the GPT frame for the story page.

This is the lesson where GPT Image 2’s detail matters most. A rooftop scene is a stack of small facts — a dove, a nail, a gun, a scar, rain on a coat — and a model that drops one of them turns a specific ending into a mood. If the first result is beautiful but empty, add the missing fact as its own clause and run it again.

Iterating Without Losing the World

Change one layer per run

Change the beat and keep the light, or the light and keep the beat. Two changes at once and you cannot tell which one broke the image.

Keep a fixed palette line

Paste the same palette clause into every prompt in a sequence: wet asphalt, warm amber, magenta signage, one cold cyan emission. Consistency is a text artifact before it is a visual one.

Keep the camera rule

If the sequence is about human smallness, every frame stays low and wide. One eye-level close-up will read as a different world.

Save the clause that worked

When a phrase fixes a problem, add it to the world card as a visual note. The next prompt starts from a better place.

Reading a Failed Generation

It looks like a photograph

Style arrived too late or detail words arrived too early. Move the painting clause to the front, cut texture adjectives, add “hard-surface planes, wet dark value groups”.

It looks like a different cyberpunk

The palette clause is missing. Add wet asphalt, warm amber and one cold cyan; remove saturated purple, green and rainbow neon.

It looks like a wall of lights

The three depth planes are missing. Name a foreground object at the bottom edge and a background silhouette at the top. Flatness is a missing plane.

It looks like a costume shop

Too many props. Keep the two that carry the beat and delete the rest. A briefcase and a coat are enough for Deckard; a dove and a nail are enough for Roy.

GPT Image 2 vs Krea 2: Which for Which Shot

City establishing shotGPT Image 2. It builds foreground, canyon and skyline as separate planes and fills the street with believable props. Krea 2 flattens the same scene into bands and loses the geography.
Character portraitGPT Image 2 for the detailed version, Krea 2 for the graphic version. Run both from the same prompt; use the GPT one to check costume logic, the Krea one for the cover.
Social media cropKrea 2. Strong centred silhouette, two or three values, readable at small size, and about five times cheaper per image.
A scene with a small crucial objectGPT Image 2. It will hold the nail, the dove, the snake scale or the scar. Krea 2 tends to simplify the object into a shape and lose the story.

Publishing and Reusing

Publish the image with the card it came from, not as a standalone picture. A viewer who can open Roy’s card, read the nail and the coat, and then see the rooftop image will understand why it looks the way it does. That link between text and image is the difference between a world and a gallery. Keep the non-affiliation note with any reuse, and do not present generated frames as official stills.

Checklist Before You Publish

One beat

Can you say what is happening in the next ten seconds? If not, the image is a portrait, not a scene.

Three planes

Does the frame have a foreground, a middle ground and a background? If the city is a wall, the answer is no.

One cold emission

Is there an artificial cyan or blue light that does not come from the sky? If everything is lit naturally, the machine is missing.

Correct scale

Is the person smaller than the architecture? Is the blimp small enough to be a blimp? Does the dove read at thumbnail size?

Two props maximum

Can you name the props without counting on one hand? If not, remove until you can.

No legible text

Check signage, screens and clothing. One readable word pulls the viewer out of the world faster than a bad hand.

Questions About This Workflow

How long should one image take?

About twenty minutes once you know the world: three minutes choosing the argument, four writing the layers, five generating on both models, five repairing one clause, three filing the result back into the card. The first image takes longer because you are learning the palette.

Do I need to prompt in English?

Both models were tested in English and the clause order assumes English weighting. If you write in another language, keep the style clause first and the exclusions last; the middle layers are more forgiving than the ends.

Should I always generate both models?

For a hero image, yes. The comparison is the cheapest quality check you have: if both models miss the same clause, the prompt is wrong; if only one misses it, the model is wrong for the job. For a quick variation, Krea 2 alone is enough.

Use the guide on the real world

The cards, the images and the comparison are all in one place. Generate your first key visual — or open the studio and pick a scene the guide does not cover.

Create From This World Explore the World Studio

Blade Runner is an independent, fan-built world in Neta Studio and is not affiliated with Alcon Entertainment, Warner Bros. or the rights holders of the film. Every character, place, event, faction and rule here is an editable world material rather than canonical text. The four key visuals on this page were generated inside this world with GPT Image 2 and Krea 2 from two identical prompts. Open the studio to read every card.