ACTIVE PROGRAM - FLAGSHIP

Train and tool AI to operate Blender and 3D environments as real workspaces - create, edit, light, and reason in the graph.

THESIS

Creative and industrial 3D work still hides behind expert UIs: Blender, DCC tools, game engines, twin editors. Copilots that only spit out tutorial prose leave the artist and engineer clicking the same buttons. Blender & 3D Agency trains models to speak the scene graph - nodes, modifiers, materials, cameras, physics - and to verify edits against the actual file.

THE PROBLEM SPACE

01Text-to-3D demos impress once; production needs iterative, reversible, named edits on real .blend / USD / glTF assets.

02Operators burn hours on repetitive modeling, layout, and lighting that an agent could do if it could see the graph.

03Existing assistants invent Blender API that does not exist, or apply ops that silently corrupt the scene.

04Spatial Intelligence without a DCC substrate stays theoretical - Blender is where we test agency in continuous space.

APPROACH

Curriculum from API literacy → constrained tool use → multi-step scene goals with visual and graph-diff checks.

Synthetic tasks inside Blender: place, align, remesh, UV, light, animate - with ground-truth before/after states.

Preference and process supervision that rewards correct intermediate ops, not just final screenshots.

Human review for destructive ops; Agent Cloud kill-switch and audit on every graph mutation.

Export paths toward engine and twin formats so the work transfers beyond one DCC.

ARCHITECTURE SKETCH

How the pieces fit

01

SCENE GRAPH INTERFACE

Typed actions over objects, collections, modifiers, and datablocks - with dry-run and undo as first-class primitives.

02

RENDER & DIFF CRITIC

Compares expected vs. actual geometry, material, and lighting; rejects hallucinated API sequences.

03

SKILL LIBRARY

Composable procedures (kitbash layout, product turntable, warehouse racking) that agents call like tools.

04

TRAINING FORGE

Headless Blender workers that generate trajectories at scale for supervised and reinforcement-style fine-tuning.

OPEN QUESTIONS

  • Where should the boundary sit between generative mesh models and procedural/graph edits?
  • How do we evaluate lighting and composition taste without collapsing into style averaging?
  • Can one agent policy transfer across Blender, engine editors, and industrial twin tools with shared spatial primitives?

WHY IT MATTERS

Design and manufacturing teams get agents that actually move vertices and cameras - with audit trails.

Training data flywheels for Spatial Intelligence grounded in real scene mutations.

A path from marketing 'AI art' to governed production tooling in creative ops.

RELATED PROGRAMS

Book a demo