Research map / Vision-language and multimodal / Cross-modal learning
TIGaussian: Disentangle Gaussians for Spatial-Awared Text-Image-3D Alignment
Jiarun Liu, Qifeng Chen, Yiru Zhao, Minghua Liu, Baorui Ma, Sheng Yang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add multi-branch 3DGS tokenizer (ablation) | UniGS-like baseline · Objaverse classification | top-1 accuracy: 33.64 -> 35.57 | Improved |
| add multi-view image fusion module (ablation) | TIGaussian w/ tokenizer + MV · Objaverse image retrieval | top-1 accuracy: 46.19 -> 53.75 | Improved |
| add 3D-text projector (ablation) | tokenizer-only model · Objaverse text retrieval | text retrieval top-1: 19.15 -> 20.80 | Improved |
| add multi-branch tokenizer to full pipeline (ablation) | MV+fusion+projector · Objaverse classification | top-1 accuracy: 37.72 -> 41.76 | Improved |
| add multi-view images (no fusion) (ablation) | tokenizer model · Objaverse image retrieval | top-1 accuracy: 41.68 -> 46.19 | Improved |
| add pretrained Uni3D point-cloud guidance (ablation) | 3DGS tokenizer · Objaverse-LVIS zero-shot | top-1 accuracy: 27.93 -> 41.76 | Improved |
| use 12 instead of 6 multi-view images (ablation) | TIGaussian · Objaverse-LVIS classification | top-1 accuracy: 41.76 -> 40.52 | Hurt |
| use 1 view vs 6 views (ablation) | TIGaussian · Objaverse-LVIS image retrieval | image retrieval top-1: 40.52 -> 52.11 | Improved |
| use ground-truth rendered views instead of generated (ablation) | TIGaussian · Objaverse-LVIS zero-shot | top-1 accuracy: 41.76 -> 41.03 | No clear effect |
| use Cap3D captions instead of InternLM (ablation) | TIGaussian · Objaverse-LVIS zero-shot | top-1 accuracy: 41.76 -> 40.12 | No clear effect |
| replace Hunyuan3D-v1-std views with lite (ablation) | TIGaussian · Objaverse-LVIS zero-shot | top-1 accuracy: 41.76 -> 39.98 | Hurt |
| scale training to Objaverse-800k | TIGaussian vs UniGS · Objaverse-LVIS zero-shot | top-1 accuracy: 48.8 -> 50.1 | Improved |
Closest papers by meaning
- Escaping Plato's Cave: Towards the Alignment of 3D and Text Latent Spaces (CVPR 2025 · similarity 0.89)
- UniGS: Unified Language-Image-3D Pretraining with Gaussian Splatting (ICLR 2025 · similarity 0.88)
- SOVGaussian: Sparse-View 3D Gaussian Splatting for Open-Vocabulary Scene Understanding (AAAI 2025 · similarity 0.88)
- Language Embedded 3D Gaussians for Open-Vocabulary Scene Understanding (CVPR 2024 · similarity 0.87)
- AlignVLM: Bridging Vision and Language Latent Spaces for Multimodal Document Understanding (NeurIPS 2025 · similarity 0.87)
- Tango3D: Towards Alignment for Global and Local 2D-3D Correspondence (NeurIPS 2026 · similarity 0.87)
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.
