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

Forum topic · 小凯 · 2026-06-21

Summary

JanusMesh is a training-free, text-driven framework for generating 3D visual illusions—single 3D meshes that display entirely different semantics from different viewing angles. Existing optimization-based approaches are slow and prone to oversaturated colors, while naive stitching methods produce geometry-inconsistent objects with visible unnatural seams and semantic leakage. The proposed method works in two stages: first, a cross-space dual-branch denoising process dynamically decodes 3D latents into voxel space for CLIP-guided directional alignment and signed distance field (SDF) blending, ensuring seamless geometric fusion; second, a view-conditioned texture synthesis module projects and aggregates view-specific 2D diffusion priors onto the fused geometry. Experiments show JanusMesh generates highly realistic dual-semantic 3D illusions in just 3–5 minutes, significantly outperforming existing methods in geometric integrity, semantic recognizability, and efficiency. Paper: arXiv 2506.15890, by Siang-Ling Zhang, Huai-Hsun Cheng, and Tsung-Ju Yang.

Paper Overview

Research Area: Computer Vision (CV) Authors: Siang-Ling Zhang, Huai-Hsun Cheng, Tsung-Ju Yang Published: 2026-06-19 arXiv: 2506.15890

Abstract

Creating 3D visual illusions—a single 3D mesh that presents completely different semantics from different viewing angles—is a fascinating but highly challenging task. Existing optimization-based methods are slow and tend to produce oversaturated colors, while simple stitching approaches fail to generate geometrically consistent objects, resulting in visible unnatural seams and semantic leakage.

This paper proposes JanusMesh, a fast, training-free, text-driven framework for 3D visual illusion generation. The method consists of two stages:

1. Cross-space dual-branch denoising: A denoising process dynamically decodes 3D latents into voxel space for CLIP-guided directional alignment and signed distance field (SDF) blending, ensuring seamless geometric fusion. 2. View-conditioned texture synthesis: A module that projects and aggregates view-specific 2D diffusion priors onto the fused geometry.

Results

Extensive experiments demonstrate that the proposed method generates highly realistic dual-semantic 3D illusions in only 3–5 minutes, significantly outperforming existing approaches in geometric integrity, semantic recognizability, and efficiency.

---

*Auto-collected on 2026-06-21*