ARTFEED — Contemporary Art Intelligence

GeoCache: Training-Free Acceleration of Multi-View Texture Diffusion

ai-technology · 2026-08-15

GeoCache introduces a novel technique for accelerating multi-view texture diffusion models without the need for training, aimed at producing high-quality 3D textures. This method tackles the significant computational burden associated with multiple evaluations of per-view denoisers in geometry-conditioned multi-view diffusion. While current training-free accelerators optimize computation across denoising phases, skipping steps in multi-view texturing disrupts cross-view interactions, compromising consistency and fidelity. GeoCache uncovers a useful redundancy: although intermediate features are specific to views, corresponding surface points share a transferable evolution in their predicted clean signals. By assessing a rotating selection of anchor views and transferring their geometry-aligned updates to other views, the method seeks to lower computational costs while preserving texture quality. The research is accessible on arXiv with the identifier 2608.13255.

Key facts

  • GeoCache is a training-free plugin for multi-view texture diffusion.
  • It accelerates multi-view texture generation by reducing computational cost.
  • It exploits geometric correspondence across views to transfer updates.
  • It evaluates a rotating subset of anchor views.
  • It transports geometry-aligned per-step updates to remaining views.
  • It addresses the issue of degraded consistency when skipping denoising steps.
  • The method is described in an arXiv paper (2608.13255).
  • The paper is a cross-type announcement.

Entities

Institutions

  • arXiv

Sources