DeepMonkey Studio: from generated scenes to editable worlds, with GLM-5.3

fatasia — Mon, 19 Oct 2026 04:35:45 +0000

DeepMonkey Studio is an editable 2D/3D world platform: scene editing, data, scripts, and AI operations belong to the same project. I maintain the project. You can try the browser demo or inspect the repository.

From Vibe Coding to Vibe World

Generating a striking scene is only the beginning. The harder question is what remains after the prompt: can you reopen the world, change an object, attach data, inspect an action, and ship it somewhere else?

That is the idea behind DeepMonkey's phrase “Vibe World.” The output should be a project you can keep working on, not just an image or a disposable preview. Its longer-term direction includes world models and AI for Science; those are directions, not a claim that the editor is already a general world model.

What belongs in the project

The editor brings four concerns together:

Presentation: 2D dashboards and 3D scenes, with animation, scripting, and plugins.
Meaning: ontology objects, relationships, and actions that connect visual objects to domain concepts.
Live behavior: data connectors, device topology, real-time data, and alarms.
Delivery: viewers and publish targets, with capability and dependency checks rather than an assumption that every backend supports everything.

A factory scene is a useful example. A machine is not only a mesh: it can be an object with properties and relationships, shown in a 3D scene and connected to a dashboard. The project includes industrial tooling around PlantLite, PPRLite, URDF robots, and virtual commissioning.

Why an ontology matters for AI actions

A chat box beside a viewport does not automatically understand the project. DeepMonkey's ontology layer describes objects, their relationships, and available actions so AI can work with project context rather than only generate text about a screenshot.

Operations are designed around permissions, previews, cancellation, and traceability. These controls matter when an action changes an existing project. Generating a suggestion and applying it are different operations.

The AI integration includes model adapters, Responses/Chat Completions support, and MCP. The model is a participant in the workflow, not a replacement for the saved project structure.

Editor and engine are separate layers

DeepEngine includes a TypeScript WebGPU path, Rust/wgpu native work, and WebAssembly, using shared scene packages. Studio also retains a Three.js/WebGL compatibility path.

That distinction matters when evaluating the demo: selecting a rendering backend is not proof of identical support across all deployment targets. Rendering, physics, import conversion, and publishing have backend-specific requirements. The repository documents the available paths.

There are independent SDK and scripting entry points as well. Developers can use Scene/Server SDKs or plugins without treating the whole editor UI as their application's API.

Try the right version

The GitHub Pages demo is a partial frontend demonstration. It does not require an account, stores edits locally, and includes an SMT production-line sample. It is not the full hosted backend service.

For a local editor, use Node.js 24 or later and run:

npx deepmonkey

For an SDK starter, use npx create-deepmonkey my-world. The repository also provides Windows, Docker, and source-based setup paths. Choose one of those when testing server-backed functionality.

Where GLM-5.3 fits

Development used multiple models. My estimated usage split is GPT-6 Astra 50%, GLM-5.3 40%, and Opus 10%. These percentages describe model usage, not measured percentages of authored code. I am not assigning individual engine or editor modules to a model without evidence.

This makes DeepMonkey a GLM-assisted development case, but not a GLM-only application. This article was prepared with AI assistance from the project's documented features.

For evaluation, start with the sample scene, inspect its objects, and distinguish local frontend editing from server-backed workflows. When reporting a problem, include the deployment mode, rendering backend, asset format, and reproduction steps in a GitHub issue.

WITHIN: turning a painting into a place you can explore with Three.js

fatasia — Mon, 19 Oct 2026 04:34:23 +0000

WITHIN is a browser-based 3D gallery where you can enter a scene instead of stopping at its frame. I am sharing it as the project's maintainer: live gallery · source · video collection.

A painting is a place, not just a texture

A digital gallery usually gives you a wall, a frame, and a zoom control. WITHIN takes a different direction: the picture becomes a place you can move through. Its 18 worlds include painting-inspired environments, natural landscapes, and a Roadster scene.

The aim is not archaeological reconstruction. It is to preserve the feeling of looking at a picture while giving the visitor control over distance, time, and viewpoint. The original painting remains a reference, not a claim that every building has been historically reconstructed.

Three ways to enter the same scene

The interaction is organized around three modes:

Guided chapters: follow a composed route, useful for a first visit or a short visual tour.
Free camera: leave the route and explore the scene's composition from another angle.
Walking: experience the world closer to human scale rather than treating it as a model on a turntable.

Some worlds add their own interactions, including boating in the cherry-blossom environment and driving, paint changes, and jumping in Roadster. Walking is an exploration mode; building collision coverage is not complete.

Motion carries the atmosphere

People and boats move through the scenes. Weather, petals, changing light, and music make time visible without turning every moment into an effect demonstration. In the Qingming-inspired world, the river, bridge, boats, and evening lighting are the focus.

For a 3D gallery, camera freedom and atmosphere have to work together. A scene that only looks convincing from one recorded viewpoint is closer to a film set than an explorable place. WITHIN therefore offers quality settings and desktop/mobile controls alongside its guided presentation.

Seven scenes provide original-painting comparisons. Screenshot and video export let visitors keep their own view rather than only sharing the project's trailer.

Run it locally

The stack is React, TypeScript, Three.js, and Vite. Rendering uses WebGL; this is not a WebGPU showcase. The site can be deployed as static files.

With Node.js 22.12 or later:

git clone https://github.com/fatasia/within.git
cd within
npm ci
npm run dev

The repository's original code uses MIT. Referenced artworks, music, and other third-party assets have their own attribution information in the project.

AI involvement

Development used GPT-6 Astra at medium/high reasoning settings. That describes the development workflow, not a browser runtime dependency. This introduction was also prepared with AI assistance from the project's documented features.

Start with the live gallery, follow a guided chapter, then switch to free exploration. For technical feedback, a scene name, browser/device, quality setting, and reproduction steps are more useful than a general performance score. Issues are welcome in the repository.

DEV Community: fatasia