Research map / 3D vision / 3D generation
VIAFormer: Voxel-Image Alignment Transformer for High-Fidelity Voxel Refinement
Tiancheng Fang, Bowen Pan, Lingxi Chen, Jiangjing Lyu, Chengfei Lv, Chaoyue Niu, Fan Wu
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 Image Index spatial grounding (ablation) | VIAFormer · Toys4k (VFM-derived, Pi3 input) | IoU: 0.4176 -> 0.4460 | Improved |
| add Image Index spatial grounding (ablation) | VIAFormer · Dora (VFM-derived, Pi3 input) | IoU: 0.4418 -> 0.4585 | Improved |
| add Correctional Flow (refine VFM latent instead of noise) (ablation) | VIAFormer · Toys4k (VFM-derived, Pi3 input) | IoU: 0.3102 -> 0.4460 | Improved |
| add Correctional Flow (refine VFM latent instead of noise) (ablation) | VIAFormer · Dora (VFM-derived, Pi3 input) | IoU: 0.2675 -> 0.4585 | Improved |
| add single-image cross-attention conditioning to geometry-only self-attention (ablation) | 24-Layer transformer · Toys4k (VFM-derived, Pi3 input) | IoU: 0.4111 -> 0.4255 | No clear effect |
| add single-image cross-attention conditioning to geometry-only self-attention (ablation) | 24-Layer transformer · Dora (VFM-derived, Pi3 input) | IoU: 0.4285 -> 0.4356 | No clear effect |
| use VGGT-conditioned 4-view features with 3-layer MLP adapter instead of image cross-attention (ablation) | 24-Layer Cross-Attn · Toys4k (VFM-derived, Pi3 input) | IoU: 0.4255 -> 0.4055 | No clear effect |
| use VGGT-conditioned 4-view features with 8-layer self-attention adapter instead of image cross-attention (ablation) | 24-Layer Cross-Attn · Toys4k (VFM-derived, Pi3 input) | IoU: 0.4255 -> 0.4076 | No clear effect |
| use Voxel-Image Union-Space with Image Index instead of standard cross-attention | VIAFormer vs 24-Layer Cross-Attn · Toys4k (synthetic corruption) | IoU: 0.2590 -> 0.8580 | Improved |
| use VIAFormer instead of Pi3 voxelization input | VFM-derived voxels · Toys4k (VFM-derived, Pi3 input) | IoU: 0.3959 -> 0.4460 | Improved |
| use VIAFormer instead of Pi3 voxelization input | VFM-derived voxels · Dora (VFM-derived, Pi3 input) | IoU: 0.4244 -> 0.4585 | Improved |
| increase number of input views beyond 3 to 4 (ablation) | VIAFormer · Toys4k (view count study) | IoU | Mixed |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 3 on most shared results.
- SDFusion — SDFusion: Multimodal 3D Shape Completion, Reconstruction, and Generation (CVPR 2023) · better on 3 of 3 shared results
- DiffComplete — DiffComplete: Diffusion-based Generative 3D Shape Completion (NeurIPS 2023) · better on 3 of 3 shared results
- PatchComplete — PatchComplete: Learning Multi-Resolution Patch Priors for 3D Shape Completion on Unseen Categories (NeurIPS 2022) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- MeshFormer : High-Quality Mesh Generation with 3D-Guided Reconstruction Model (NeurIPS 2024 · similarity 0.87)
- Learning with Unreliability: Fast Few-shot Voxel Radiance Fields with Relative Geometric Consistency (CVPR 2024 · similarity 0.85)
- ShapeShifter: 3D Variations Using Multiscale and Sparse Point-Voxel Diffusion (CVPR 2025 · similarity 0.85)
- SILSA: Sliding-Window Slice Latents for Topology-Preserving High-Resolution 3D Generation (NeurIPS 2026 · similarity 0.85)
- GeoSVR: Taming Sparse Voxels for Geometrically Accurate Surface Reconstruction (NeurIPS 2025 · similarity 0.85)
- CVT-xRF: Contrastive In-Voxel Transformer for 3D Consistent Radiance Fields from Sparse Inputs (CVPR 2024 · similarity 0.84)
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.
