Abstract
Generating high-fidelity 3D point clouds requires balancing global structural coherence with fine-grained local details. Traditional diffusion models often employ static architectures that may not be optimal for all stages of the denoising process. We present TIGER (Time-Varying Denoising Model), a novel framework that dynamically adapts its processing strategy. By utilizing a two-stream architecture—a Transformer for global shape and a CNN for local refinement—coupled with a learnable "time mask," TIGER achieves state-of-the-art results on ShapeNetv2 with significantly improved training efficiency.
Introduction: The Challenge of 3D Point Cloud Generation
Generating realistic 3D point clouds is a complex task. A model must simultaneously capture the overall "gist" or topology of an object (e.g., the general shape of a chair) and the nuanced geometry of its surfaces (e.g., the smoothness of the seat).
Current state-of-the-art diffusion models often rely on static architectures, such as UNet-style convolutional designs. However, the requirements of the denoising process change over time. At high noise levels (early stages), the model must establish global structure. At low noise levels (late stages), the focus shifts to refining local details and smoothing surfaces. Static CNNs often lack the receptive field for global structure, while standard Transformers can be computationally prohibitive and may miss local consistency.
Methodology: Time-Varying Two-Stream Architecture
The core of TIGER is its Time-Varying Two-Stream Architecture, which decouples global and local feature extraction to optimize the denoising trajectory.
The Two Streams
- Transformer Branch: This branch captures long-range dependencies to establish the overall object shape. It incorporates a novel Position-Aware Self-Attention (PASA) mechanism.
- CNN Branch: A shallow Point-Voxel CNN (PVCNN) is used to model local relationships and refine surface quality, ensuring geometric consistency.
Dynamic Feature Fusion
A key innovation is the Time Mask Generator. This module takes the current diffusion timestep $t$ as input and produces an optimizable mask. This mask dynamically reweighs the contributions of the Transformer (global) and CNN (local) features. As noise decreases, the contribution of the convolutional branch increases monotonically, allowing the model to "sand down" surfaces once the global "sketch" is complete.
Continuous Position Encoding
To enhance the Transformer's spatial awareness, TIGER introduces two continuous position encoding methods: Phase Shift Position Encoding (PSPE) and Base-$\lambda$ Position Encoding (B$\lambda$PE). Unlike discrete alternatives, these methods represent 3D coordinates in a continuous space, preserving linear relationships between relative positions.
Latent Denoising
To maintain computational efficiency, TIGER operates on a downsampled "latent" point cloud. A specialized encoder-decoder structure facilitates the transition between the full point cloud and this efficient latent representation.
Experimental Results
TIGER was evaluated primarily on the ShapeNetv2 dataset, focusing on categories like Airplane, Chair, and Car, as well as a "universal" model trained on all 55 categories.
Performance and Efficiency
- State-of-the-Art Metrics: TIGER outperformed leading models such as PVD and LION in Chamfer Distance (CD) and Earth Mover’s Distance (EMD). Notably, it achieved a 13.45% improvement in 1-NN accuracy over PVD for the airplane category.
- Training Speed: TIGER demonstrated a significant efficiency advantage, reaching peak performance in approximately 164 GPU hours, compared to 550 GPU hours for LION—a 3.3x speedup.
- Visual Fidelity: Qualitative analysis shows that TIGER produces point clouds with fewer artifacts and more coherent global structures, such as straight airplane wings and well-defined chair legs, compared to baselines like PointFlow or SetVAE.
Conclusion and Future Work
TIGER demonstrates that a time-varying architecture is highly effective for 3D diffusion processes. By dynamically balancing global and local processing, the model achieves superior quality with less computational cost. Future research may explore unifying these two streams into a single network with inherent time-varying properties or applying this framework to tasks like point cloud completion and upsampling.
Summary updated