Research map / Point clouds / Shape and surface completion

TIGER: Time-Varying Denoising Model for 3D Point Cloud Generation with Diffusion Process

CVPR 2024 · Point clouds › Shape and surface completion · approach: Point cloud diffusion generation · 14 citations (OpenAlex)

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 TIGER, a time-varying two-stream diffusion denoiser that fuses a shallow CNN branch and a latent point Transformer with a learned timestep-dependent mask for 3D point cloud generation.

Problem

Diffusion models for 3D point cloud generation typically use UNet/PVCNN convolutional denoisers, which struggle to capture global shape early in denoising; the architecture's role across timesteps is understudied.

Approach

Combine convolution (local detail) and Transformer attention (global shape) in a two-stream denoiser whose fusion weights are a learned function of the diffusion timestep, plus two continuous 3D position encodings (PSPE and BλPE) and position-aware self-attention.

Key results

• On airplanes, 1-NN accuracy with EMD improves to 56.26 vs 61.23 for LION under LION's split (TIGER reports outperforming LION in four of six metrics).
• Relative to PVD, airplane 1-NN (CD) improves 13.45% (73.82 to 71.85 under the standard split reported in Table 1) and 8.12% relative to LION.
• Training takes 164 GPU hours vs 550 for LION (about a quarter), with inference 9.73 s vs 27.12 s on V100 GPUs.

Measured effects reported in the paper

Surprising findings

• The learned ConvNet weight increases nearly monotonically with timestep, i.e., convolution matters more at later denoising stages.
• Standard Transformer attention beats convolution on EMD while convolution is better on CD, suggesting complementary metric behavior.

Limitations

• Cannot control the category of generated shapes (unconditional only).
• Two sub-networks (CNN and Transformer) for local and global features make the model less efficient.

Experimental setup

Datasets: ShapeNetv2

Benchmarks: ShapeNetv2 airplane ShapeNetv2 chair ShapeNetv2 car

Metrics: 1-NN accuracy (CD) 1-NN accuracy (EMD) training GPU hours inference time

Baselines: r-GAN l-GAN PointFlow DPF-Net ShapeGF SoftFlow SetVAE DPM PVD LION

Builds on: denoising diffusion probabilistic models PVCNN PVD LION Point Transformer

Closest papers by meaning

• Where Precision Meets Efficiency: Transformation Diffusion Model for Point Cloud Registration (AAAI 2025 · similarity 0.88)
• 3DMambaIPF: A State Space Model for Iterative Point Cloud Filtering via Differentiable Rendering (AAAI 2025 · similarity 0.87)
• SuperPC: A Single Diffusion Model for Point Cloud Completion, Upsampling, Denoising, and Colorization (CVPR 2025 · similarity 0.87)
• Denoising Diffusion via Image-Based Rendering (ICLR 2024 · similarity 0.87)
• TexOct: Generating Textures of 3D Models with Octree-based Diffusion (CVPR 2024 · similarity 0.87)
• Scaling Diffusion Models to Real-World 3D LiDAR Scene Completion (CVPR 2024 · similarity 0.86)