Source-linked AI summary
CreativeFlow: A One-to-Many Analogical Relation Transfer Method for 3D Asset Generation
Xuechen Li, Shuai Zhang, Nanxuan Zhao, Qing Chen
TL;DR
Text-to-3D systems can produce homogenized, semantically conservative results for open-ended prompts, with limited structural variation. CreativeFlow uses analogical relations to generate diverse targets from a source asset, and expert evaluations report higher visual-fascination ratings than the strongest baseline alongside identity-preservation and novelty scores.
Problem
Open-ended text-to-3D prompts can yield homogenized, semantically conservative outputs with shallow style or color variations rather than novel structural deviations.
Source: §1 · pp. 1–2Method
Results
3.31/5 versus 2.84 for the strongest baseline in visual fascination; experts also rated CreativeFlow 3.66 for identity preservation and 3.86 for novelty.
Source: §3 · p. 2Takeaways & Limitations
The workflow and resulting assets are proposed as a foundation for a structured source-relation-target dataset and benchmark for future relation-aware 3D model training.
Source: §4 · p. 2
Abstract
from arXiv · showhide
Inspired by cognitive science, we present CREATIVEFLOW, an analogical generation framework that explicitly models analogical divergent thinking to mitigate creative homogenization in text-to-3D pipelines. Our method derives a series of meaningful yet relationally similar source-target asset pairs, each featuring distinct geometric configurations. Expert evaluations demonstrate that our framework substantially enhances creative novelty and visual fascination. This workflow and its resulting assets establish a foundational dataset and benchmark for future relation-aware 3D model training.
1 Introduction
Current text-to-3D systems can collapse open-ended prompts into semantically conservative forms and shallow style or color variations. CreativeFlow instead uses analogical relational mapping to transfer meaningful structural, functional, or material logic and generate diverse target designs.
1 Introduction
- Existing text-to-3D systems often collapse open-ended creative prompts into semantically conservative forms and shallow style or color variations.They tend to explore few novel structural deviations of the source object [Li et al. 2026].Source: §1 · pp. 1–2
- Figure 2 illustrates transfers of aesthetic, structural/material, and mixed relations from source assets to diverse yet recognizable targets.Source: Figure 2 · p. 2
- Analogical creativity frames prompts as relational mappings from a source object’s attributes or structures to a target design through an analogical concept.The formulation aims to transfer structural, functional, or material logic rather than superficial details.Source: §1 · p. 2
- CreativeFlow maps a single source object to diverse, relation-aware targets, using knowledge-graph retrieval and LLM associations to expand candidate relations.Candidates are filtered for semantic novelty, 3D feasibility, and structural transferability before being translated into instructions for 2D-prior-to-3D generation.Source: §1 · p. 2
2 Methodology
CreativeFlow addresses convergent text-to-3D outputs by mapping a source asset through inferred attributes, dual-path relational expansion, and analogical target generation. The pipeline combines structured knowledge-graph retrieval with LLM-internal associations, then filters and reconstructs target designs.
2 Methodology
- Existing models converge on dominant semantic modes because they do not explicitly model transferable relations between local components and cross-domain concepts.Source: §2 · p. 2
- CreativeFlow first uses an LLM to infer a source asset’s functional, material, structural, visual, and affordance attributes as transferable seed terms.The seeds can include parts such as a support frame, seat shell, or backrest.Source: §2 · p. 2
- Dual-path expansion finds analogical sources through knowledge-graph queries and LLM-internal transfer of historical cases.The knowledge-graph path grounds sources and attributes in Wikidata, then uses Getty AAT for aesthetic attributes and AskNature for biological structures and biomimetic mechanisms.Source: §2 · p. 2
- The pipeline solves an analogy between source attributes and a target attribute, filters mappings for novelty, feasibility, and transferability, then generates 2D priors and reconstructs 3D meshes.For example, chair support-frame structure can be mapped to coral branching geometry.Source: §2 · p. 2
3 Preliminary Evaluation and Results
CreativeFlow produces multiple geometrically distinct targets from a source asset by exploring aesthetic and structural relations. In expert studies, it received favorable ratings for identity preservation, novelty, and visual fascination relative to the evaluated baselines.
3 Preliminary Evaluation and Results
- A single source asset expands into 5 to 10 aesthetic and structural relations, each yielding targets with diverse geometric deformations.Source: §3 · p. 2
- 3.31/5 versus 2.84 for the strongest baseline in visual fascination; CreativeFlow also scored 3.66 for identity preservation and 3.86 for novelty.Two expert user studies compared matched-prompt assets against Hunyuan3D-2, TRELLIS-2, and SF3D using five-point scales.Source: §3 · p. 2
4 Future Work
The paper identifies future work in creativity evaluation, relation-aware training data, and human-computer creative tools. The figures provide examples of the pipeline and one-to-many qualitative outputs.
4 Future Work
- The authors plan to develop standardized metrics for generative creativity and analogical structural rationality.Source: §4 · p. 2
- They plan a large-scale structured source-relation-target dataset for training and benchmarking future models [Deitke et al. 2023].Source: §4 · p. 2
- They aim to integrate the analogical expansion pipeline into collaborative design tools for intuitive, controllable, and inspiring human-AI ideation.Source: §4 · p. 2
- Figure 3 depicts source-attribute inference, dual-path concept expansion, and target-attribute reconstruction; Figure 4 shows one-to-many outputs for stools, handbags, and frogs.Source: Figure 3 · p. 3, Figure 4 · p. 3