AI3D Foundation / Research
The interface should translate—not the human.
Humans express ideas through many representations: drawings, words, images, rough geometry, objects, gestures, and spatial arrangements.
AI systems often reverse that relationship. They require a person to translate an idea into whichever representation a particular model happens to accept.
AI3D researches another possibility.
We build representations and interfaces that let people work in the forms most natural to them, while the system translates those expressions into whatever representations different models require—and translates the results back into forms that remain visible, spatial, and editable.
3D is especially useful within this loop because it can persist across transformations.
It can be input, prompt, constraint, composition, intermediate representation, output, environment, and context for intelligence.
Canonical works
Each system asks a version of the same question: What translation is the human being forced to perform that the interface could perform instead?
You should not have to enter the computer’s authoring environment before computation can begin.
Napkinmatic treats ordinary drawings and physical artifacts as computational interfaces. Work can begin with the representation already in front of the person rather than inside a specialized tool.
Computation moves toward the person’s existing expression.
You should not have to describe a drawing in another modality before it can become spatial.
A napkin, notebook, receipt, or other available surface becomes a direct bridge from human visual intuition to generated geometry.
Natural expression crosses into a spatial representation.
You should not have to leave the creative environment every time you switch models or representations.
AI3D Desktop combines text-to-3D, image-to-3D, 3D-to-3D, and direct spatial manipulation inside a persistent skeuomorphic XR environment. Generation becomes one operation within an ongoing workspace.
Multiple representations begin to coexist in one continuing environment.
You should not have to translate shape into prose or a privileged 2D projection.
A rough 3D construction communicates proportion, orientation, topology, perspective, and spatial relationships directly. Shape itself becomes the prompt for generative 3D.
Spatial intent becomes a first-class input rather than only an output.
You should not have to describe spatial composition indirectly to image or video generators.
Scenes, cameras, composition, and animation become controls for generative image and video. Generated media returns to an iterative loop in which rough geometry and model output can inform the next creative action.
3D mediates between spatial intention and models built for other media.
You should not have to know or care which representation a particular model requires.
A person can sketch, trace, place, stretch, compose, or manipulate forms directly in space. Canvas maps that expression into the representation a model needs, then maps the result back into an editable 3D world.
Representation translation becomes infrastructure beneath the creative environment.
Once translation disappears beneath the interface, intelligence can operate on the shared world itself rather than waiting for isolated prompts.
Spatial Creation Agents receive the existing world—including its geometry, objects, relationships, and composition—as context. They can create in relation to what is already there.
The system moves from translating between a human and a model to allowing both to operate within the same continuing environment.
Earlier Work
A selection of earlier research, spatial computing systems, interactive media, and experimental interfaces is being reconstructed for the archive.
Hackathons
Selected hackathon projects, prototypes, and award-winning experiments will be documented separately.