A purpose-built game studio for people & AI.
SceneFoundry is a bespoke studio for building a hidden-object game — its scenes, the objects hidden inside them, the maps, tokens and sound. AI didn’t just make the assets; it made it viable to build the tool at all — one that works in the flow of the work, instead of swivel-chairing between Photoshop, an audio suite and a spreadsheet, and writing custom bridges to carry data between them.
And it keeps a person at the centre on purpose. Whether a scene looks right, whether an object is fun to hunt, whether a puzzle lands as easy or unfair — none of that is verifiable by an AI today. It requires taste. So a person directs, the machine generates, and the person decides.
A person and AI working together to achieve what neither could make independently.
Draw the blueprint once. Every scene inherits it.
Before a single scene exists, the author writes the adventure’s blueprint — the world it’s set in, the style it’s drawn in, and the voice its objects speak in. That adventure-level meta-prompt is the throughline: it’s what keeps twenty-three scenes of one world looking, and reading, like one world. People bring the taste and the intent; the models hold the line.
- WorldAn authoring guide the models treat as ground truth — never shown to players.
- StyleA flat-2D-illustrated switch, or photoreal — set per adventure.
- VoiceDescriptions grounded in real-world fact, or purely in this world’s lore.
You describe it. The model drafts it. You choose.
A scene isn’t typed into a prompt box and accepted. It’s a short back-and-forth — the human sets direction, the model does the heavy lifting, and the human keeps the final say at every gate.
That loop — direct → generate → judge — repeats for every object, cutout, description and sound in the world. The AI never ships anything a person didn’t choose.
Find an object, and it has to leave something behind.
The game ships each scene as a clean background plate with object cutouts composited on top. Tap a hidden object and its cutout fades away — revealing the plate beneath. So the tool has to know what was always behind every findable, and paint it. That inpainted patch is the quiet, load-bearing problem SceneFoundry exists to solve — at hundreds of objects per adventure.
Drag the handle and watch it happen: the Weaver Bird lifted out, the empty nest seamlessly filled — then a person confirming the reveal actually looks right. That last step is the whole argument in miniature. The fill is generated; whether it reads as the branch that was always there is a judgement.
The model’s first pass, a person’s finishing tools.
Infill is one move in a much larger toolkit. Every extracted object opens into a full editor: a prompt you can rewrite and regenerate, alongside pixel tools to clean the cutout by hand — no round-trip to Photoshop. The AI gets you 90% of the way; the human tools close the last 10% that makes an asset actually shippable.
And the words, in bulk — then curated.
Display names, lore descriptions and thumbnails generate as mass actions across every object at once — then a human reads down the review table and approves, edits, or regenerates each one. Volume from the machine, judgement from the person.
No single model is best at everything — so none is hard-wired.
Each job routes to the provider that’s currently best at it, through one facade and a queued lane per model. Swapping is a per-request decision, not a rewrite.
| Job | Provider | Model |
|---|---|---|
| Segmentation | Replicate | yodagg / sam3-image-seg |
| Image generation | Gemini OpenAI | gemini-3-pro-image · gpt-image-2 |
| Inpaint & removal | Gemini OpenAI | gemini-3.1-flash-image · gpt-image-2 |
| Descriptions & metadata | Gemini | gemini-3-flash-preview |
| Sound & ambience | ElevenLabs | eleven_text_to_sound_v2 |
| Enhance / upscale | Krea | Reimagine |
An adventure builder — end to end.
SceneFoundry doesn’t stop at scenes. One tool generates the illustrated world map and places its locations, forges the achievement tokens and title graphics, writes and generates the ambient sound, and holds every scene with its objects. A single author takes an adventure from blank to shippable.
Not a demo — an instrumented tool in daily use.
SceneFoundry meters itself — not for a dashboard, but to find out where the work actually hurts. The action counts say which motion an author repeats hundreds of times a week, and that is where the next hour of tool-building goes. Friction you can see is friction you can remove; the tool improves in the places the telemetry says it should.
The authoring tool is larger than the game it feeds — its backend alone is two-thirds the size of the whole web game, and its biggest single file is the infill engine, sceneInpaint.ts, at 2,387 lines. (Count the whole product — web plus the native port — and the two are roughly on par.)
That’s the tell. When the tool to make the content is as large as the content engine itself, the moat isn’t the game — it’s owning the pipeline where one person’s taste and a rack of models make more, and better, than either could alone.