GPT Image 2 OMORI World Generator — Two Worlds in One Square Frame
A working GPT Image 2 generator for OMORI-style worlds: the three-pass prompt, the crayon-watercolour style contract, both literal prompts, and an honest comparison with Krea 2 on the same two briefs.
GPT Image 2 is a general image model that becomes an OMORI world generator when you give it a world instead of a mood. Ask for "OMORI art" and you get a pale boy in a striped shirt standing in a field. Give it a specific country — a train station, a picnic blanket, a pink castle on a hill, an impossible white rectangle in the grass — plus a style contract and a detail budget, and you get a frame that looks like it came from the game's press kit. This page is the working manual for that difference, with both literal prompts published in full.
Everything below was generated inside OMORI — Headspace and Faraway Town: four published works, two prompts, two models (GPT Image 2 and Krea 2), the same 1024×1024 size, no reference images and no retouching. The world supplies the canon and the style contract; the model supplies the decisions. The comparison is kept honest by changing nothing else.
Why a World Beats a Franchise Name
A franchise name is a retrieval cue, not a description. Say "OMORI" and the model reaches for the most repeated images of the game: a monochrome boy, a red hand, a light bulb. Those are icons, not scenes. A world prompt replaces the icon with geography, weather, inhabitants and one wrong object. The model can draw a place it has been told about; it can only average a place it has heard of.
| Setting | What the prompt says | What the model does |
|---|---|---|
| Prompt | Two named children, one porch, one basketball | Renders a scene with behaviour and distance |
| Style | Crayon line, watercolour wash, reserved white | Keeps the medium consistent across both prompts |
| Geography | Station left, meadow centre, castle on the hill | Builds a readable depth stack instead of a collage |
| Detail budget | One eye-scribble, one white rectangle, one bike | Places the odd detail instead of filling every space |
| Negative list | No text, no letters, no logo, no watermark | Avoids the poster-typography reflex |
| Aspect | 1024×1024, square | Both models can match it exactly, so the comparison is fair |
The Three-Pass Prompt
Pass one — geography
Name the frame and the depth bands: station at the left edge, meadow in the centre, castle on the horizon. Say what is in each band. Without this the model invents a collage.
Pass two — weather and one light
Late summer, soft diffuse light, paper white reserved for the brightest edge. One direction and one temperature. Two light sources will make the watercolour look like a filter.
Pass three — the wrong object
A blank white rectangle cut into the meadow; an eye-scribble in the trees; a bike left where it fell. Exactly one impossible thing. Two is a poster; one is a story.
Pass four — style contract
Medium, line quality, edges, palette, light model, detail budget. Keep it under 80 words and reuse it verbatim on every prompt in the world.
Pass five — negative list
No text, no letters, no logo, no watermark, no UI, no realistic shading, no extra characters. The list is short because every item is a real failure you have seen.
The order matters more than the wording. Geography first forces the model to commit to a layout before it starts decorating. Style second locks the medium. The wrong object last gives the model permission to be strange inside a frame it already understands.
The OMORI Style Contract
This world uses one style contract for every prompt. It is deliberately written as instructions about materials, not adjectives about mood. "Handmade" is not a material; "waxy broken crayon outlines with reserved paper white" is.
Medium: mixed crayon and transparent watercolor illustration. Line: waxy crayon outlines that break, overlap and leave paper grain. Wash: transparent overlapping watercolor layers; edges may diffuse naturally. White: paper white reserved for the brightest light, never painted over. Colour: waxy brights with pastel washes; local colour stays identifiable. Light: soft diffuse light built from washes, no realistic light-shadow steps. Detail: large rounded simplified shapes; detail budget of one odd object per frame. Forbid: photography, PBR, heavy digital paint, comic inking, text, logos, watermarks, extra characters.
Reuse it word for word. The moment you paraphrase the contract between prompts, the two images stop belonging to the same world, and the comparison stops meaning anything.
Prompt One — Headspace, Two Worlds, One Boy
This is the establishing frame. It has to do four things at once: prove the dream is beautiful, prove it is handmade, plant the exit, and keep the cast to a single figure so the landscape can breathe. The white rectangle is the exit. The eye-scribble is the price of using it.
Hand-drawn crayon and transparent watercolor illustration, square composition. A single wide view of Headspace: a bright handmade country of patchwork crayon-green forest, a red-and-white picnic blanket with a wicker basket, a small wooden playground, a candy-pink castle far on the horizon, and a little train station platform at the left edge. In the foreground a pale, expressionless boy in black-and-white clothes stands with his back to the viewer at the edge of a blank white rectangle of paper that cuts unnaturally into the meadow, like a doorway into an empty room. A black scribble with a single white eye coils at the treeline behind him. Waxy broken crayon outlines, reserved white paper, transparent overlapping watercolor washes, pastel brights, soft diffuse light, no realistic shading, no text, no letters, no logo, no watermark.
What GPT Image 2 Did With It
GPT Image 2 treated the prompt as a painting brief. It built a long soft afternoon, put grain in the paper, gave the station a roof with individual tiles and the castle a flag, and rendered Omori as a boy with a specific posture rather than a symbol. It also invented a line of flowers along the bottom edge and a small sign on the station — details that were not asked for, but that do not contradict the world. The eye-scribble became a fine ink tangle in the upper-right trees, present but not shouting.
| Criterion | GPT Image 2 | Krea 2 |
|---|---|---|
| Composition | Painterly depth, readable path | Centred poster, flatter bands |
| Omori | Painted boy, back turned, specific | White oval head, two dots, black shorts |
| Something | Fine ink tangle in the canopy | Bold eye and thick coil on the lawn |
| Castle | Pink towers, texture, flag | Blocky pink silhouette, crenellations |
| Station | Detailed roof, bench, sign | Small house shape at the frame edge |
| Picnic | Red check, basket, scattered flowers | Red check, basket, no flowers |
| Light | Directional, soft, bloom in the corners | Even, decorative, almost shadowless |
| Thumbnail read | Atmospheric, needs a beat to parse | Instant, icon-like, very legible |
| Best use | Hero image, landing page, press | Poster, social card, icon, iteration |
Prompt Two — Kel Knocks, the Porch
The second prompt moves to Faraway Town and changes the social temperature. It is the same world and the same style contract, but the frame is now a doorway with two people on opposite sides of glass. The impossible object is not a white rectangle; it is the boy in the dark hallway who should be visible and is not.
Hand-drawn crayon and transparent watercolor illustration, square composition. A sunlit front porch of a small white clapboard house in late summer: a boy in an orange hoodie holds a basketball under one arm and knocks on a closed screen door. Behind the glass is a dark hallway where a pale, silent boy stands just out of the light. Overgrown lawn, a bicycle on the path, cicada heat in the air, one wooden step worn smooth. Waxy broken crayon outlines, transparent watercolor washes, reserved paper white for the harsh sunlight, soft diffuse shadows, handmade childlike rendering, quiet and warm with something wrong underneath. No text, no letters, no logo, no watermark.
What Changes on Krea 2, and Why It Is Useful
Krea 2 is not a worse GPT Image 2; it is a different reading of the same brief. It compresses the scene into shapes that survive a 200-pixel thumbnail. That is not what you want for a hero image, and it is exactly what you want for a social card, an icon, a map pin or a quick iteration before you spend a slower model on the final.
- Icon legibility Krea 2's Omori is a white oval and two dots. At 64 pixels it is still Omori. GPT Image 2's painted boy becomes a smudge.
- Flat colour Krea 2 treats the wash as a fill. The result is closer to a picture book than a watercolour painting, which is a legitimate OMORI reading.
- Symmetry Krea 2 centres the porch automatically. Use it when you want a formal, stage-like composition; use GPT Image 2 when you want a camera.
- Speed and cost Krea 2 is the cheaper iteration model. Generate four compositions, pick the one that reads, then re-run the winner on GPT Image 2 at high quality.
- Text weakness Krea 2 is known to struggle with legible text. Since every prompt ends with a text ban, this rarely matters here — but never ask it for a title.
Parameters That Actually Matter
- 1024×1024 The comparison size. Both models support it, so no difference in the images can be blamed on the aspect ratio.
- 1536×1024 GPT Image 2 only. Better for a wide hero or a landing-page banner; Krea 2 cannot match the aspect.
- quality=high Worth it for a key visual that will be inspected. Skip it while you are still deciding on a composition.
- quality=medium The iteration setting. Generate four, pick one, then re-run the winner at high.
- No reference images This comparison deliberately uses text only. Adding a reference changes the variable being tested.
- One prompt, two models The cheapest way to learn a model's taste. Two images teach more than ten prompts on one model.
Do not change size and model at the same time. If the two images differ in aspect, the eye reads the crop as a style difference and the comparison becomes noise.
Failure Modes Worth Knowing
The Iteration Loop, in Four Moves
Write the world
Read the 25 atoms first. The prompt is downstream of the canon: the station, the picnic, the castle and the eye-scribble all come from the world page, not from memory.
Draft on the cheap model
Run the composition on Krea 2 at 1024×1024. You are checking layout, not finish. Two or three drafts is usually enough to find the frame.
Finish on the detail model
Re-run the chosen composition on GPT Image 2 at quality=high. Keep the prompt identical. The only variable should be the model.
Publish both, with captions
Keep the pair. Publish the prompts. The comparison is the content, and the captions are where you admit what the model invented.
The loop is deliberately boring. Most bad AI art comes from changing five things at once and then guessing which one helped. One variable at a time is slower for the first image and much faster by the fourth.
Using These Images in a World Page
- Hero and open graph Use the GPT Image 2 Headspace frame. It has depth, a readable subject and enough empty light on the left for a title overlay.
- Scene reference Use the porch pair when a roleplay scene needs a shared picture. The Krea 2 version is better for a small inline card; the GPT Image 2 version for a full-width plate.
- Style guide Put the style contract next to the images. Anyone adding a new location should be able to copy the block and stay inside the world.
- Attribution Credit the model, the size and the fact that no reference images were used. The transparency is part of the artefact.
Generate a Headspace you can walk into
Open the world, copy the style contract and run the three-pass prompt. Start with the establishing frame, then the porch, then the pair on the second model. Both prompts on this page are copy-paste ready.
GPT Image 2 + Krea 2, one world
Two prompts, two models, four published works — the exact prompts are on this page.
OMORI — Headspace and Faraway Town is a fan-built world on Neta Studio. OMORI and its characters belong to OMOCAT. All four images were generated for this world at 1024×1024 with GPT Image 2 and Krea 2, without reference images or post-processing. Prompts and comparison notes are published so the process can be checked.