Fast3Dcache: Training-free 3D Geometry Synthesis Acceleration
Mengyu Yang, Yanming Yang, Chenyi Xu, Chenxi Song, Yufan Zuo, Tong Zhao, Ruibo Li, Chi Zhang
2025-12-01
Summary
This paper focuses on making the process of creating 3D shapes with diffusion models much faster, without sacrificing the quality of the resulting shapes.
What's the problem?
Diffusion models are really good at generating things like images and 3D models, but they take a long time to actually *create* something because they work by gradually refining a result through many steps. Speeding things up by reusing calculations from previous steps, which works well for 2D images, doesn't work as well for 3D because even tiny errors in the reused data can build up and cause noticeable flaws and distortions in the final 3D shape.
What's the solution?
The researchers developed a new system called Fast3Dcache that cleverly reuses calculations in a way that’s designed for 3D shapes. It predicts which parts of the 3D model are stable and won't be affected by reusing data, and then focuses on reusing calculations for those areas. They use two main ideas: a way to figure out how much data can be reused based on how stable the shape is, and a way to pick the most reliable data to reuse based on how quickly it's changing.
Why it matters?
This work is important because it significantly speeds up the creation of 3D models using diffusion models – up to 27% faster with a big reduction in the amount of computing power needed – while still maintaining a high level of detail and accuracy in the generated shapes. This could make it much more practical to use these powerful models for applications like design, gaming, and virtual reality.
Abstract
Diffusion models have achieved impressive generative quality across modalities like 2D images, videos, and 3D shapes, but their inference remains computationally expensive due to the iterative denoising process. While recent caching-based methods effectively reuse redundant computations to speed up 2D and video generation, directly applying these techniques to 3D diffusion models can severely disrupt geometric consistency. In 3D synthesis, even minor numerical errors in cached latent features accumulate, causing structural artifacts and topological inconsistencies. To overcome this limitation, we propose Fast3Dcache, a training-free geometry-aware caching framework that accelerates 3D diffusion inference while preserving geometric fidelity. Our method introduces a Predictive Caching Scheduler Constraint (PCSC) to dynamically determine cache quotas according to voxel stabilization patterns and a Spatiotemporal Stability Criterion (SSC) to select stable features for reuse based on velocity magnitude and acceleration criterion. Comprehensive experiments show that Fast3Dcache accelerates inference significantly, achieving up to a 27.12% speed-up and a 54.8% reduction in FLOPs, with minimal degradation in geometric quality as measured by Chamfer Distance (2.48%) and F-Score (1.95%).