Andrii Bidochko

• Updated: June 15, 2026
• 6 min read

CubePart: An Open-Vocabulary Part-Controllable 3D Generator

Direct Answer

CubePart introduces an open‑vocabulary, part‑controllable 3D generation framework that lets developers specify a custom list of semantic parts (e.g., “wing”, “tail”, “handle”) and receive a coherent set of meshes that assemble into a complete object. This capability bridges the gap between monolithic generative models and the modular asset pipelines required by modern games, simulations, and robotics.

Background: Why This Problem Is Hard

Interactive 3D assets in games, VR, and autonomous‑robot simulations are rarely single, undivided meshes. Instead, they are composed of semantically meaningful parts that can be animated, swapped, or physically simulated independently. Traditional generative 3D models—such as implicit‑field networks or diffusion‑based mesh generators—produce either:

• A single, monolithic mesh that must be manually segmented before it can be rigged or assigned physics properties.
• An arbitrary part decomposition that does not align with the designer’s intent (e.g., a “chair” split into random patches rather than “seat”, “back”, “legs”).

These limitations force developers to spend hours on post‑processing, undermining the promise of AI‑driven content creation. Moreover, existing part‑aware pipelines rely on closed vocabularies (fixed part names) or require extensive annotation, making them brittle when new object categories emerge.

What the Researchers Propose

CubePart proposes a two‑stage generative architecture that decouples global shape synthesis from part‑level decoding, while exposing an open‑vocabulary part schema as an explicit control signal at inference time. The framework consists of three logical components:

1. Global Shape Encoder‑Decoder: Generates a coarse, holistic latent representation of the target object based on a free‑form text prompt (e.g., “a futuristic hoverboard”).
2. Part Schema Interpreter: Takes the user‑provided list of part names—unrestricted in length or terminology—and maps each term to a semantic embedding that conditions the downstream decoder.
3. Part‑Level Mesh Decoder: For every schema element, a dedicated decoder produces a high‑resolution mesh that respects both the global shape context and the local part semantics.

By separating the concerns of overall geometry and fine‑grained part detail, CubePart can generate meshes that are both globally consistent and semantically aligned with arbitrary, open‑vocabulary part specifications.

How It Works in Practice

The end‑to‑end workflow can be visualized as a pipeline of four stages:

1. Prompt & Schema Input: The user supplies a natural‑language description of the object and a comma‑separated list of desired parts (e.g., “handle, blade, guard”).
2. Global Latent Generation: A transformer‑based text‑to‑latent model converts the prompt into a compact shape code that captures the overall silhouette and style.
3. Part Conditioning: Each part name is embedded using a large language model’s vocabulary, then fused with the global code via cross‑attention. This step ensures that the “blade” inherits the sword’s overall curvature while retaining its own semantic identity.
4. Mesh Decoding & Assembly: Independent decoders output per‑part meshes. A lightweight alignment module translates, rotates, and snaps the parts together based on learned spatial priors, yielding a ready‑to‑use asset bundle.

The following illustration shows the high‑level architecture and data flow:

Key differentiators of CubePart include:

• Open‑vocabulary control: No predefined taxonomy; any textual part label can be used.
• Scalable data pipeline: A semi‑automated process harvested millions of part‑labeled meshes from public repositories, enriching them with language annotations.
• Two‑stage generation: Global coherence is enforced before any part is rendered, reducing inter‑part collisions and ensuring physical plausibility.

Evaluation & Results

CubePart was benchmarked across three representative scenarios:

1. Semantic Fidelity Test: Human evaluators rated how well generated parts matched the supplied names. Over 92% of parts were deemed “accurate” or “highly accurate,” a substantial jump from the 58% baseline of monolithic generators.
2. Assembly Consistency Check: Automated collision detection measured inter‑part intersections. CubePart achieved a 0.04 average intersection‑over‑union score, compared to 0.21 for a state‑of‑the‑art diffusion model.
3. Game‑Engine Integration Demo: Assets were exported to Unity and Unreal Engine without additional retopology. Animators successfully rigged a “robotic arm” generated with parts “base, joint, gripper,” confirming immediate usability.

Beyond raw numbers, the experiments demonstrated that developers can now iterate on part vocabularies (adding or removing parts) and receive instantly updated meshes, dramatically shortening the asset creation loop.

Why This Matters for AI Systems and Agents

For AI‑driven pipelines that assemble virtual worlds, the ability to request and receive modular meshes on demand unlocks several strategic advantages:

• Dynamic Content Generation: Game AI agents can request new objects with specific functional parts (e.g., “a door with a lock and a peephole”) during runtime, enabling procedurally generated quests.
• Simulation Fidelity: Robotics simulators can request objects whose parts map directly to physical parameters (mass, hinge constraints), improving the realism of training environments.
• Cross‑Domain Reuse: The same part schema can be shared between design tools, AI planners, and behavior scripts, fostering a unified asset language.

Integrating CubePart into an AI orchestration platform such as the UBOS platform overview allows developers to chain the generator with downstream modules—like physics engines or voice‑controlled agents—through a single API call. This reduces engineering overhead and accelerates time‑to‑market for AI‑enhanced interactive experiences.

What Comes Next

While CubePart marks a significant step forward, several open challenges remain:

• Fine‑Grained Control: Current part conditioning captures high‑level semantics but lacks explicit control over sub‑features (e.g., “curved blade vs. straight blade”). Future work could integrate attribute‑level prompts.
• Scalability to Complex Assemblies: Generating objects with dozens of parts (e.g., a car) may strain the decoder’s capacity. Hierarchical decomposition strategies are a promising direction.
• Cross‑Modal Consistency: Aligning generated meshes with textures, materials, and sound effects in a single pass remains an unsolved problem.

Addressing these gaps will broaden the applicability of part‑controllable generation to domains such as digital twins, e‑commerce product configurators, and AI‑assisted industrial design. Companies looking to embed such capabilities into their workflows can explore the AI marketing agents offering, which already leverages modular content generation for personalized campaign assets.

References

For a complete technical description, see the original CubePart paper.

Andrii Bidochko

CTO UBOS

Andrii Bidochko is an AI entrepreneur and researcher focused on AI agents, reinforcement learning, and autonomous systems. He writes about the technologies shaping the future of machine intelligence, from frontier models and agent architectures to real-world AI applications.