Generate a 3D product mockup from a one line prompt with editable parts
Iterate on a single component of a model without regenerating the whole shape
Swap between Claude, GPT, and Gemini to compare 3D code generation quality
Run a local browser based 3D viewport for previewing AI generated GLB files
Needs Flutter 3.24 plus an OpenAI, Anthropic, or Gemini API key, and you must use 127.0.0.1 instead of localhost or Google sign in breaks.
Nova3D is a tool for making 3D objects from a text prompt or a reference image. The thing that sets it apart, according to its README, is that the parts of the object stay separate and editable instead of being fused into a single shape. For example, you could generate a washing machine and then change the door or a knob without having to redo the whole model. The trick behind this is the way Nova3D produces 3D content. Most AI 3D tools generate a single mesh directly from images. Nova3D instead writes Python scripts that drive Blender, the open-source 3D program. The result is a GLB file with named parts arranged in a clear hierarchy, including proper materials and textures. The README contrasts this with diffusion-based tools that output a merged blob, and with code-based CAD systems that struggle with organic shapes. This repository is the client app, written in Dart with Flutter. It talks to a hosted service at nova3d.xyz that is currently closed source. To run it locally you need Flutter 3.24 or higher, then you clone the repo, run flutter pub get, and start it as a local web server on port 5555. You open the page in a browser, sign in with Google or email, paste in an API key for OpenAI, Anthropic, or Gemini, and type a prompt. The client is model agnostic, so you can switch between Claude, GPT, and Gemini in the settings. It also includes a built-in 3D viewport based on Three.js with transform tools and material editing, local caching of past generations in the browser, and support for reference images that guide the generated script. The README notes a couple of gotchas, mainly that you must use the 127.0.0.1 address rather than localhost for sign-in to work. The project is open source on the client side, but self-hosting the backend would mean writing your own server.
Generated 2026-05-22 · Model: sonnet-4-6 · Verify against the repo before relying on details.