Research map / 3D vision / 3D generation
TeMO: Towards Text-Driven 3D Stylization for Multi-Object Meshes
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes TeMO, a text-driven 3D stylization framework for multi-object meshes that uses a cross-modal graph attention module (DGA) to align noun phrases with object surface points and a cross-grained CLIP contrast loss (CGC) combining coarse sentence-level and fine word-level supervision.
Problem
Text-driven 3D stylization methods built on CLIP work well for single objects but fail on multi-object meshes, mixing object properties and losing local details due to coarse global image-text contrast.
Approach
Decouple noun phrases from the text and cluster mesh surface points per object, link matching phrase and point nodes in a cross-modal graph for cross-attention (DGA), and add a word-level fine-grained CLIP contrast loss alongside the global one (CGC).
Key results
- TeMO reaches an alignment score of 0.285 versus 0.274 for TANGO and 0.265 for X-Mesh, and 0.262 for Text2Mesh.
- In a 60-user study, TeMO scores 3.344 / 3.311 / 3.261 on naturalness, text match, and style match, versus 2.406 / 2.450 / 2.539 for TANGO.
- Ablation on the two-dragon scene shows the DGA module distinguishes the two objects and the CGC loss adds semantic detail (e.g., fire vs ice), with both together giving the best result.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TeMO instead of TANGO | Multi-object 3D stylization · Multi-object scenes | Alignment: 0.274 -> 0.285 | Helped |
| Use TeMO instead of TANGO | Multi-object 3D stylization · Multi-object scenes | User-Q1 naturalness: 2.406 -> 3.344 | Helped |
| Use TeMO instead of TANGO | Multi-object 3D stylization · Multi-object scenes | User-Q2 content match: 2.450 -> 3.311 | Helped |
| Use TeMO instead of TANGO | Multi-object 3D stylization · Multi-object scenes | User-Q3 target-style match: 2.539 -> 3.261 | Helped |
| Add the DGA module (ablation) | TANGO baseline · Two dragons; “a fire dragon and an ice dragon” | Object distinction | Helped |
| Add the CGC loss (ablation) | TANGO baseline · Two dragons; “a fire dragon and an ice dragon” | Semantic detail | Helped |
| Add both the DGA module and CGC loss (ablation) | TANGO baseline · Two dragons; “a fire dragon and an ice dragon” | Stylization quality | Helped |
Surprising findings
- The 3D symmetry prior commonly used in prior work causes interference between objects in multi-object scenes, so the authors removed it for all compared methods.
Limitations
- Removing the 3D symmetry prior is a limitation; symmetry-prior-based consistency for each object is not yet incorporated.
- Diffusion models struggle to generate multi-object images, so diffusion-based stylization is not applicable to multi-object scenes.
- Relies on NLTK noun-phrase extraction and GMM clustering of surface points, which add pipeline complexity.
Experimental setup
Datasets: COSEG Thingi10K ShapeNet TurboSquid ModelNet
Benchmarks: multi-object 3D mesh stylization test scenes (authors' collection)
Metrics: CLIP score (alignment) user study mean opinion score (naturalness, text match, styl
Baselines: Text2Mesh TANGO X-Mesh CLIP-Mesh
Builds on: TANGO CLIP X-Mesh Text2Mesh TACo
Closest papers by meaning
- Towards Text-guided 3D Scene Composition (CVPR 2024 · similarity 0.87)
- Multi-StyleGS: Stylized Gaussian Splatting with Multiple Styles (AAAI 2025 · similarity 0.86)
- GraphDreamer: Compositional 3D Scene Synthesis from Scene Graphs (CVPR 2024 · similarity 0.86)
- CLIP-Driven Open-Vocabulary 3D Scene Graph Generation via Cross-Modality Contrastive Learning (CVPR 2024 · similarity 0.86)
- Styl3R: Instant 3D Stylized Reconstruction for Arbitrary Scenes and Styles (NeurIPS 2025 · similarity 0.86)
- MARVEL-40M+: Multi-Level Visual Elaboration for High-Fidelity Text-to-3D Content Creation (CVPR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.