Research map / 3D vision / 3D generation
3DOT: Texture Transfer for 3DGS Objects from a Single Reference Image
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 3DOT, a training-based framework that transfers texture from a single reference image onto a 3D Gaussian Splatting object using progressive view propagation, view-consistency gradient guidance, and prompt-tuning-based gradient guidance.
Problem
Transferring texture from a single 2D reference image onto a 3DGS object while keeping view consistency across unseen views and preserving reference texture characteristics.
Approach
Progressively edits views starting from the reference view and propagates outward, guides diffusion with gradient terms derived from differences between reference-conditioned and unconditioned diffusion branches, and learns a CLIP-aligned token capturing texture difference between the unedited and reference renderings.
Key results
- Ours achieves CLIP score 0.9333 vs. 0.8917 for the best baseline (IN2N) and LPIPS (Alex) 0.1166 vs. 0.1683 (IGS2GS).
- Vision-GPT score 76 vs. 54 (DGE) and user study 4.575 vs. 2.4375 (IGS2GS).
- Editing takes 23min 33s, versus 15min 47s for GaussCtrl and 20h 51min for IN2N.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use 3DOT instead of the strongest baseline, IN2N | 3D texture transfer · Evaluation set | CLIP Score: 0.8917 -> 0.9333 | Helped |
| use 3DOT instead of the strongest baseline, IGS2GS | 3D texture transfer · Evaluation set | Lpips(Alex): 0.1683 -> 0.1166 | Helped |
| use 3DOT instead of the strongest baseline, GaussCtrl | 3D texture transfer · Evaluation set | Lpips(VGG): 0.1591 -> 0.1247 | Helped |
| use 3DOT instead of the strongest baseline, DGE | 3D texture transfer · Evaluation set | Vision-GPT: 54 -> 76 | Helped |
| use 3DOT instead of the strongest baseline, IGS2GS | 3D texture transfer · Evaluation set | User study: 2.4375 -> 4.5750 | Helped |
| add prompt-tuning-based gradient guidance (ablation) | 3DOT · Texture transfer ablation | LPIPS (Alex): 0.0355 -> 0.0351 | Helped |
| add prompt-tuning-based gradient guidance (ablation) | 3DOT · Texture transfer ablation | LPIPS (VGGT): 0.0679 -> 0.0620 | Helped |
| add prompt-tuning-based gradient guidance (ablation) | 3DOT · Texture transfer ablation | CLIP Score: 0.9203 -> 0.9445 | Helped |
| add view-consistency gradient guidance (ablation) | 3DOT · Texture transfer ablation | LPIPS (Alex): 0.0558 -> 0.0351 | Helped |
| add view-consistency gradient guidance (ablation) | 3DOT · Texture transfer ablation | LPIPS (VGGT): 0.0844 -> 0.0620 | Helped |
| add view-consistency gradient guidance (ablation) | 3DOT · Texture transfer ablation | CLIP Score: 0.9340 -> 0.9445 | Helped |
| enable progressive generation (ablation) | 3DOT · Texture transfer ablation | LPIPS (Alex): 0.1330 -> 0.0351 | Helped |
Limitations
- Dark borders around edited regions caused by imperfect language-based segmentation masks.
- Quality depends on the unedited 3DGS; undertrained floaters can disrupt mask generation and cause geometry collapse.
- Significantly different reference geometry can degrade transfer quality.
Experimental setup
Datasets: 360-degree scenes face-forwarding scenes
Benchmarks: 360-degree dataset face-forwarding dataset
Metrics: CLIP score LPIPS (AlexNet) LPIPS (VGG) Vision-GPT score user study
Baselines: IN2N IGS2GS GaussCtrl DGE Plug-and-Play
Builds on: Instruct-GS2GS GaussCtrl Direct Gaussian Editor Textual Inversion classifier-free guidance
Closest papers by meaning
- GuideFlow3D: Optimization-Guided Rectified Flow For Appearance Transfer (NeurIPS 2025 · similarity 0.87)
- Textured Gaussians for Enhanced 3D Scene Appearance Modeling (CVPR 2025 · similarity 0.86)
- ReGS: Reference-based Controllable Scene Stylization with Gaussian Splatting (NeurIPS 2024 · similarity 0.86)
- TriTex: Learning Texture from a Single Mesh via Triplane Semantic Features (CVPR 2025 · similarity 0.86)
- One-2-3-45++: Fast Single Image to 3D Objects with Consistent Multi-View Generation and 3D Diffusion (CVPR 2024 · similarity 0.86)
- Text-Guided 3D Face Synthesis - From Generation to Editing (CVPR 2024 · similarity 0.86)
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.