JanusMesh: Fast and Zero-Shot 3D Visual Illusion Generation via Cross-Space Denoising

2026-06-18

AI Summary

Existing methods for generating multi-view 3D visual illusions with semantically distinct appearances often suffer from slow inference, color oversaturation, or geometric inconsistency, leading to visible seams and semantic leakage. This work proposes a training-free, two-stage framework that first achieves geometrically seamless modeling by performing cross-space dual-branch denoising in voxel space, integrating CLIP-guided directional alignment with signed distance field (SDF) fusion. Subsequently, a view-conditioned texture synthesis module projects and aggregates view-specific 2D diffusion priors onto the fused geometry. The proposed approach is the first to enable fast (3–5 minutes), zero-shot generation of high-quality dual-semantic 3D illusions, significantly outperforming existing methods in geometric completeness, multi-view semantic clarity, and computational efficiency.

Abstract

Creating 3D visual illusions, a single 3D mesh that reveals entirely different semantics from various viewing angles, is a fascinating but tough challenge. Existing optimization-based methods are slow and can produce oversaturated colors. In contrast, naive stitching approaches fail to produce geometrically coherent objects. This results in visible unnatural seams and semantic leaks. In this paper, we present a fast and training-free framework for generating text-driven 3D visual illusions. Our approach decouples the generation into two stages. First, we propose a cross-space dual-branch denoising process. This process dynamically decodes 3D latents into voxel space for CLIP-guided orientation alignment and Signed Distance Field (SDF) blending, which ensures seamless geometric fusion. Second, we introduce a view-conditioned texture synthesis module that projects and aggregates view-specific 2D diffusion priors onto the fused geometry. Extensive experiments demonstrate that our method generates highly realistic, dual-semantic 3D illusions in just 3-5 minutes. It significantly outperforms existing methods in geometric integrity, semantic recognizability, and efficiency. Project page: https://siang1105.github.io/JanusMesh.github.io/

Problem

• 3D visual illusion
• zero-shot generation
• geometric coherence
• semantic leakage
• view-dependent semantics

Innovation

• cross-space denoising
• 3D visual illusion
• zero-shot generation
• Signed Distance Field (SDF)
• view-conditioned texture synthesis