SIGGRAPH Asia 2026

CreativeFlow: A One-to-Many Analogical Relation Transfer Method for 3D Asset Generation

Xuechen Li, Shuai Zhang, Nanxuan Zhao, Qing Chen

Tongji University · Adobe Research

Paper ↗ Read HTML ↗ DOI ↗

One source. Many creative directions.

Open-ended text-to-3D prompts often lead to similar shapes. CreativeFlow uses analogical divergent thinking to turn a single source asset into multiple designs with distinct geometry. Explicit relations guide the transfer of structure, function, material, and aesthetics while keeping the source recognizable.

Comparison of standard 3D generation, CreativeFlow outputs, and human design references — From homogenized outputs to structurally diverse designs: CreativeFlow transfers meaningful relations through structure mapping.

How CreativeFlow works

The framework models a proportional analogy: A : A′ :: B : B′. A source asset and its attributes are connected to an analogical concept and the attributes that can be transferred back into a new design.

Attribute inference

An LLM identifies transferable parts and attributes across function, material, structure, visuals, and affordances.

Dual-path expansion

Knowledge graph retrieval uses Wikidata, Getty AAT, and AskNature. A second path draws on the LLM’s internal associations to expand possible analogical concepts.

Mapping & generation

Candidate mappings are filtered for semantic novelty, 3D feasibility, and structural transferability. Part-level instructions guide 2D visual priors and 3D reconstruction.

CreativeFlow pipeline: attribute inference, dual-path expansion, and analogical mapping with 3D generation — The three-stage pipeline converts a source prototype into diverse relation-guided 3D assets.

Relation-guided design exploration

Each source expands into 5–10 aesthetic and structural relations. For example, relating a chair’s support frame to coral’s branching structure transfers a concrete geometric principle into the target design.

Examples of aesthetic, structural and material, and mixed analogical relations — Aesthetic, structural/material, and mixed relations produce diverse yet recognizable target designs.

One-to-many generated results for stools, handbags, and frogs with multi-view renderings — One-to-many results across stools, handbags, and frogs. Each row connects a source asset to generated targets and their multiple views.

Preliminary evaluation

Two expert studies compared matched prompts against Hunyuan3D-2, TRELLIS-2, and SF3D using five-point scales.

3.66 / 5

Identity preservation

Mean score for retaining a recognizable source identity.

3.86 / 5

Novelty

Mean score for creative novelty in the identity-preservation and novelty study.

3.31 / 5

Visual fascination

Ten design experts rated CreativeFlow at 3.31, compared with 2.84 for the strongest baseline.

These preliminary findings support the framework’s potential for novel, recognizable, and visually engaging 3D design.

Future directions

Future work will develop evaluation metrics for generative creativity and analogical structural rationality, build a large-scale source–relation–target dataset, and explore interactive tools for collaborative human–AI ideation.

Content and figures adapted from CreativeFlow, arXiv:2610.05167v1 (CC BY 4.0).