Research map / Point clouds / Shape and surface completion
PointInfinity: Resolution-Invariant Point Diffusion Models
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 PointInfinity, a two-stream transformer point-cloud diffusion model with a fixed-size latent that trains at low resolution (1k points) and generates up to 131k points at test time, improving surface fidelity on CO3D.
Problem
Point cloud diffusion models are limited to low resolutions (2k-4k points) because transformer cost scales quadratically with the number of points, and vanilla transformers are not resolution-invariant when tested at different resolutions.
Approach
Decouple a fixed-size latent representation for the underlying surface from a variable-size point stream using read/write cross-attention, so compute grows linearly with output points and the model can be trained at low resolution and sampled at high resolution; higher test resolution improves fidelity, analogous to a fidelity-variability tradeoff like classifier-free guidance.
Key results
- Test-time resolution scaling from 1k to 8k points lowers CD@1k from 0.227 to 0.181 and raises PSNR from 13.37 to 14.27 on CO3D-v2.
- At 16k test resolution, PointInfinity reaches CD@1k 0.179 and FS 0.724 versus Point-E 0.197/0.675 and MCC 0.234/0.549.
- Generation scales to 131k points, 31x more than Point-E, with inference time and memory growing linearly in test resolution.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| increase test-time resolution from 1024 to 8192 | PointInfinity · CO3D-v2 | CD@1k: 0.227 -> 0.181 | Helped |
| increase test-time resolution from 1024 to 8192 | Point-E · CO3D-v2 | CD@1k: 0.239 -> 0.232 | Hurt |
| use PointInfinity instead of Point-E | CO3D-v2 generation at test resolution 16k · CO3D-v2 | CD@1k: 0.197 -> 0.179 | Helped |
| use PointInfinity instead of MCC | CO3D-v2 generation at test resolution 16k · CO3D-v2 | FS: 0.549 -> 0.724 | Helped |
| train at resolution 64 instead of the default resolution 1024 (ablation) | PointInfinity · CO3D-v2, test resolution 16k | CD@1k: 0.179 -> 0.178 | Helped |
| train at resolution 2048 instead of the default resolution 1024 (ablation) | PointInfinity · CO3D-v2, test resolution 16k | CD@1k: 0.179 -> 0.183 | Hurt |
| use 64 rather than 256 latent tokens (ablation) | PointInfinity · CO3D-v2, test resolution 16k | CD@1k: 0.179 -> 0.457 | Hurt |
| use 128 rather than 256 latent tokens (ablation) | PointInfinity · CO3D-v2, test resolution 16k | CD@1k: 0.179 -> 0.182 | Hurt |
| use 512 rather than 256 latent tokens (ablation) | PointInfinity · CO3D-v2, test resolution 16k | CD@1k: 0.179 -> 0.176 | Helped |
| increase mixture baseline resolution from 1024 to 8192 points (ablation) | Mixture baseline · CO3D-v2 | CD@1k: 0.227 -> 0.211 | Helped |
| use PointInfinity instead of the mixture baseline at 8192 points (ablation) | CO3D-v2 · CO3D-v2 | CD@1k: 0.211 -> 0.181 | Helped |
| restrict the read module to 1024 input tokens (ablation) | PointInfinity · CO3D-v2, test resolution 8192 | CD@1k: 0.181 -> 0.224 | Hurt |
Surprising findings
- Higher test-time resolution slightly improves surface quality even though the model was trained at 1024 points, and Point-E degrades at higher resolution.
- Restricting the read module to training-size tokens removes the improvement, supporting the hypothesis that more input tokens carry more information between denoising steps.
Limitations
- Point-E was reimplemented because official training code was not released, so baseline numbers may differ from the original.
- Resolution invariance is empirical rather than strict mathematical invariance.
- Evaluation is limited to CO3D-v2, with noisy COLMAP-derived point clouds.
Experimental setup
Datasets: CO3D-v2
Benchmarks: CO3D-v2
Metrics: Chamfer Distance F-score PSNR variability (pairwise CD)
Baselines: MCC Point-E PVD ShapeGF Shap-E PerceiverIO
Builds on: DDPM Perceiver IO Point-E MCC classifier-free guidance
Closest papers by meaning
- Invaria: Learning Scale and Density Invariance in Point Clouds via Next-Resolution Prediction (NeurIPS 2026 · similarity 0.89)
- SuperPC: A Single Diffusion Model for Point Cloud Completion, Upsampling, Denoising, and Colorization (CVPR 2025 · similarity 0.88)
- Point Cloud Pre-training with Diffusion Models (CVPR 2024 · similarity 0.88)
- Where Precision Meets Efficiency: Transformation Diffusion Model for Point Cloud Registration (AAAI 2025 · similarity 0.87)
- PCDreamer: Point Cloud Completion Through Multi-view Diffusion Priors (CVPR 2025 · similarity 0.87)
- Adaptive 3D Reconstruction via Diffusion Priors and Forward Curvature-Matching Likelihood Updates (NeurIPS 2025 · 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.