Everything the studio does, written down.

Getting started

The studio is two programs: a hub that serves the interface and keeps the library, and a worker that does the generating. They can run on one machine or on two. The hub is plain Node with no dependencies; the worker is Python with the open models it needs.

# the hub (port 3000)
node hub/server.js

# the worker, on the machine that will do the work
python worker/worker.py

Open http://localhost:3000, type what you want, choose a mode, press Generate. The asset appears in the Library tab when it is done, with its files, its quality report and a record of exactly what made it.

Hosting the hub

The hub is one Node file with two optional packages, so it runs anywhere Node runs. On Replit, import the zip, set the WORKER_KEY secret, press Run; publish it as a Reserved VM and give it an App Storage bucket so the library survives redeploys. The worker stays on your own machine and connects outward to the hub's address. The hub also serves this site and the Studio, so one address is the whole product.

The four modes

Every mode answers the same prompt. They differ in how long they take and how much is generated versus reused. The engine picks the route inside each mode; you pick how long you can wait.

Times were measured on an ordinary laptop with an integrated GPU and no graphics card. A consumer graphics card is projected to bring Standard to one or two minutes; that figure is a projection until it is measured.

While it builds

In Quick, Standard and Studio the viewer shows a build card: the steps of the plan, the one the worker is on, its latest line, how long it has been and how long it usually takes. Nothing is shown as a result until the AI result exists. Behind the scenes the fast engine still makes a rough stand-in the moment you press Generate, so the record has a shape to link to; it is never shown as the answer and it stays out of the Library. If a job fails, the card says so, offers Try again, and lets you look at the stand-in if you want to.

The same words in the same mode, or in a lighter one, reuse the finished asset instead of building it again; the status line says same words already built once · reused when that happens. Tick Fresh or press Reroll to build again, or pick a more thorough mode.

Engines

The Engine control normally stays on auto. The others exist for when you know what you want: neural forces the picture-and-shape route, fast the parametric generator, library an open-licence library match, human the anatomical base for people, personal your own uploaded collection, databank the Instant path, and reskin keeps a shape the studio already has and paints it with a fresh picture.

Writing prompts

Say one thing. The studio reads the subject first, then colour, material, size and style words, and it builds a picture and a shape of that one subject. Two subjects in one sentence become one confused object.

• Subject first. "a red fox", "a wooden treasure chest with iron bands", "a samurai helmet with a gold crest".
• Colour and material words are honoured. "a black fox" gives black fur; "a bronze dragon statue" gives bronze. Words that are not in the sentence are not invented: a fox stays fox-coloured unless you say otherwise.
• Size in metres if it matters. "0.5 m" sets the real size written into the file; otherwise the studio uses the subject's typical size.
• Real animals and objects get a description from the encyclopedia so the picture is drawn from what the thing actually looks like, not from a guess.
• Revolved things are exact. Vases, barrels, bottles, goblets and columns are turned from their own silhouette, so their geometry is clean by construction.

What does not work well yet: crowds, scenes, text on surfaces, and thin structures like wire fences. Say so in the prompt and the studio will still try, but expect Standard or Studio and a look at the result.

Pictures as input

Attach a photo or drawing with the prompt. By default the studio turns the picture into 3D: the subject is cut out of any background, its shape is solved from the picture, and the picture itself paints the model. What is shown is what you get: a whole figure gives a whole model, a head-and-shoulders portrait gives a bust. No drawing step runs, so even Quick mode spends its time on a full-detail shape and a super-resolved texture.

Choose borrow its look instead when you want a different framing or pose: the studio draws a fresh picture in the same colours and style from your words (say "the whole figure, standing") and builds from that. Use it for busy photos too.

Photos from open sources are recorded with their author and licence in the asset's record; the record travels with the asset.

What comes out

• GLB with PBR materials: albedo, roughness and metalness maps, a baked normal map with tangents, real UV charts from an industry unwrapper. Every file passes the official glTF validator with zero errors before you see it.
• OBJ and MTL alongside, for tools that want them.
• Three levels of detail in Standard and Studio: about 24,000, 8,000 and 2,000 triangles, each with its own charts and textures.
• Textures at 2048 px for the full model, JPEG for colour and roughness, PNG for the normal map.
• A record of the prompt, the mode, the models and licences involved, the quality report, the judge's score and the time it took.

The quality gate

Every Standard and Studio asset is checked before it is saved: welded topology, hairline cracks stitched and small holes closed before the surface is unwrapped, floating pieces removed, winding fixed, stance checked so the thing stands on its feet, UV coverage and density measured, materials complete, and a perceptual check of the model against its own picture. Small problems are repaired in place. A severe failure sends the job back for another attempt with a different seed before anything ships.

Hardware

Licences and commercial use

Every part of the studio is open, and every asset records which parts made it. You may use what you make commercially, with one condition that comes from the shape model's licence.

This page is a summary, not legal advice. The licence texts ship with the studio.

Hand-off and API

This site hands prompts to the studio through the address bar: /?prompt=a bronze dragon statue opens the studio with the prompt filled in. The hub also takes requests directly.

POST /api/gen3d
{ "prompt": "a grey wolf", "mode": "standard",
  "style": "realistic", "quality": "medium",
  "size": "auto", "engine": "auto" }

→ { "id": "9cf5557a57c0", ... }

GET /api/gen3d/9cf5557a57c0/model.glb      the asset
GET /api/gen3d/9cf5557a57c0/model-lod1.glb the 8k level
GET /api/gen3d/library                     everything made so far

Workers register with a key and poll the hub for jobs; nothing on the hub needs an open port towards the worker.

FAQ

Is it really free?

During the beta, yes: every mode, every file. The work happens on your machine or on machines you point at the studio, so the only cost today is electricity. Access during the beta is for people who follow @Plasmated3D on X: the first time you open the studio it asks for the follow, opens X in a new tab, and remembers in that browser. Pricing after the beta will be announced before it applies.

Why does Instant look approximate?

Because it is a fit of something already made, recoloured to your words, and it says so. Use it to sketch; use Standard or Studio for the final.

Can the studio learn from what it makes?

Yes. Finished animals join a shared-topology class; the next animal of that class can be fitted in seconds rather than generated. The class grows with every good result.

Where do the files live?

In the hub's data folder, one directory per asset, with the record beside the model. Nothing leaves your machine unless you send it.