Frame claims 100 million polygons at 60 fps in a browser-based 3D editor

Ask AI to build a part, open the result in a browser and adjust its dimensions manually: that is the workflow proposed by Frame's creator. He is developing a 3D editor with its own graphics engine and an API for agents. For developers of engineering applications, two aspects are particularly interesting: how it handles demanding geometry and how a human and AI edit the same scene.

The headline figure in the publication is 100 million polygons at 60 frames per second without instancing, meaning without repeatedly rendering copies of the same geometry. The demo caption reports 76 FPS on an RTX 2080. These are the creator's results: the supplied material includes neither independent testing nor a detailed comparison methodology.

How Frame handles demanding geometry

Frame runs on a custom WebGPU engine without Three.js. According to the developer's description, geometry stays on the GPU, intensive calculations run in shaders, and some operations use WASM. Moving an entire object changes its transformation matrix; there is no need to rewrite the coordinate arrays for all its vertices.

Imagine an imported STL model of a complex part. The user rotates it, moves the camera closer and repositions the object. In this workflow, what matters is which data the editor transfers between the CPU and GPU and which work it repeats with each action. The creator attributes Frame's speed to the architecture of the entire editor, designed around WebGPU.

The claimed size of the core after gzip compression is around 0.93 MB in the measured build. Plugins and library textures load as needed. This separation benefits the product: someone doing ordinary modeling should not have to wait for every engineering module to load. The figure applies specifically to the core, however; the total download size depends on the features selected.

The editor also offers spline surfaces, Boolean operations and path-traced rendering. But the speed at which a scene rotates does not, by itself, show how quickly topology changes, complex materials or other operations run. The creator's claim of superiority over Cinema 4D should be treated as a reason to reproduce the comparison with your own workloads.

The agent gets access to the same scene

Frame has its own AI API. According to its creator, the agent accesses the same objects and operations available to a human user. The result it builds can be edited manually and saved, and its actions can be undone.

For example, the agent prepares a part's geometry, after which the user moves elements using conventional tools. The shared scene makes it possible to continue working without transferring the result to a separate editor. The project includes Doc, a text description that helps convey the context of the collaborative work.

A separate plugin, Atomic, calculates atomic interactions on the GPU through WebGPU. The creator specifies the NEP89-1 neural-network interatomic potential and support for 75 elements. The intended workflow is to have the agent prepare the structure and initial conditions, then adjust forces and constraints and repeat the experiment.

Here, ease of setting up a task must be distinguished from the reliability of the calculation. The ability to assemble a scene using words does not establish that the selected potential is suitable for a particular material and set of conditions. The supplied text contains no detailed validation of Atomic, so it is too early to draw conclusions about the accuracy of its engineering predictions.

Ideas to apply in your own development

The creator says he began writing Frame manually, then handed a substantial share of the implementation to AI. He retained responsibility for the architecture, interface, algorithms and verification of the results. Among the practices he finds useful are a detailed application description in app.md, automated tests and conventional scripts for tasks that do not need a model.

This is the experience of a single project, without a measured comparison of team productivity before and after adopting AI. Still, it offers concrete ideas to test: document how the application works, give the agent operations on actual working objects, and preserve the ability to inspect and undo the result.

According to its creator, the official version of Frame is free to use, but the code is proprietary. Compatibility between versions is not yet guaranteed. A slicer, structural analysis, thermal analysis and other engineering modules are still in development.

A practical next step is to test Frame on a copy of your own scene: import, edit, save and reopen it. For a product team, the most useful question here is how easily a user can inspect and correct an agent's work in a familiar interface.