JanusMesh Generates Multi-View 3D Illusions Without Training

JanusMesh produces 3D meshes that show different semantics from different viewing angles using a two-stage, training-free pipeline. A cross-space dual-branch denoising process decodes 3D latents into voxel space for CLIP-guided orientation alignment and SDF blending, avoiding the seams and semantic leaks of naive stitching while being faster than optimization-based methods.