Architecture · Published 26 August 2026
Why AI 3D should generate a scene graph, not one opaque mesh
The architecture behind named objects, bounded tool calls, editable hierarchy, revisions, and recoverable browser state.
A scene is a system of relationships
A useful 3D world contains objects, groups, transforms, materials, lights, cameras, constraints, and semantic zones. Compressing that into one generated mesh removes the handles a director needs for the next instruction.
Yugma keeps a structured graph so “move the forklift,” “warm the loading-bay lights,” and “duplicate the rack row” remain explicit mutations.
The model proposes tools, not bytes
Director output is decoded as typed SSE events and validated against a bounded tool catalog. The model cannot write arbitrary scene JSON or invoke an unknown operation.
Each accepted mutation identifies its target and expected revision before persistence.
Revisions preserve intent under collaboration
Owner-scoped revision checks stop an offline tab or collaborator from replacing a newer scene. The offline queue replays only compatible work and surfaces a conflict instead of merging blindly.
Recovery boundaries isolate panel failures while the durable scene remains mounted.
OpenUSD is a continuation boundary
OpenUSD-aware assets and exports preserve hierarchy and naming so downstream Omniverse or DCC workflows can continue. Compatibility must be tested per asset and schema; a file extension alone is not proof.
The production contract is therefore graph first, validation at every mutation, and explicit export qualification.
Production checklist
- Verify Revision safety against the deployed environment, not a screenshot.
- Verify Owner isolation against the deployed environment, not a screenshot.
- Verify Strict streaming against the deployed environment, not a screenshot.
- Verify State-preserving recovery against the deployed environment, not a screenshot.