Large 3D reconstruction models have revolutionized the 3D content generation field, enabling broad applications in virtual reality and gaming. Just like other large models, large 3D reconstruction models suffer from hallucinations as well, introducing structural outliers (e.g., odd holes or protrusions) that deviate from the input data. However, unlike other large models, hallucinations in large 3D reconstruction models remain severely underexplored, leading to malformed 3D-printed objects or insufficient immersion in virtual scenes. Such hallucinations majorly originate from that existing methods reconstruct 3D content from sparsely generated multi-view images which suffer from large viewpoint gaps and discontinuities. To mitigate hallucinations by eliminating the outliers, we propose Dehallu3D for 3D mesh generation. Our key idea is to design a balanced multi-view continuity constraint to enforce smooth transitions across dense intermediate viewpoints, while avoiding over-smoothing that could erase sharp geometric features. Therefore, Dehallu3D employs a plug-and-play optimization module with two key constraints: (i) adjacent consistency to ensure geometric continuity across views, and (ii) adaptive smoothness to retain fine details.We further propose the Outlier Risk Measure (ORM) metric to quantify geometric fidelity in 3D generation from the perspective of outliers. Extensive experiments show that Dehallu3D achieves high-fidelity 3D generation by effectively preserving structural details while removing hallucinated outliers.
Figures & tables
Figure 1: Our Dehallu3D generates high-quality and high-fidelity 3D meshes, effectively mitigating mesh outliers while adaptively preserving sharp features. The red and blue boxes highlight noticeable outlier regions in the Baseline method, contrasted with the corresponding regions in Dehallu3D (Ours) to showcase improvements.
Figure 2: Overview of Dehallu3D. Dehallu3D first generate orthographic multi-view color images and corresponding normal maps, which are used to initialize a coarse mesh. Next, a globally plausible mesh is quickly constructed through the Coarse Mesh Reconstruction stage. Finally, the proposed Cyclic View Consistency Refinement (CVCR) module is employed to mitigate outliers and further refine the mesh.
Appearance metrics
Geometry metrics
Methods
PSNR ↑
SSIM ↑
LPIPS ↓
Clip-Sim ↑
CD ↓
F-Score ↑
Wonder3D [ 15 ]
20.4963
0.8908
0.1851
0.6970
0.02183
0.3580
TripoSR [ 31 ]
20.5309
0.8897
0.1841
0.7146
0.02241
0.3847
InstantMesh [ 40 ]
20.8954
0.8903
0.1749
0.7538
0.02198
0.4046
CRM [ 34 ]
21.1265
0.8889
0.1720
0.7191
0.02163
0.3967
Unique3D [ 37 ]
20.9795
0.8882
0.1742
0.7493
0.02175
0.4073
Table 1: Quantitative comparison results. We mark the best scores in bold and the second-best scores with an underline .
Figure 3: Qualitative comparison of different methods in mesh reconstruction. The red and blue boxes mark specific regions in the meshes of other methods for comparison with the corresponding optimized regions in the mesh generated by Dehallu3D, while red circles highlight defects in meshes.
Figure 4: Qualitative comparison results for ablation study on LDC and LDS in CVCR module.
Appearance metrics
Geometry metrics
LSE
LDC
LDS
PSNR ↑
SSIM ↑
LPIPS ↓
Clip-Sim ↑
CD ↓
F-Score ↑
✗
✗
✗
16.8972
0.8474
0.2283
0.6315
0.02572
0.3201
✔
✗
✗
17.6534
0.8615
0.2102
0.6678
0.02449
0.3468
✗
✔
✗
20.9276
0.8823
0.1697
0.7392
0.02217
0.3879
✗
✗
✔
18.3124
0.8749
0.2019
0.6834
0.02332
0.3545
✗
✔
✔
21.1973
0.8894
0.1531
0.7403
0.02114
0.4097
Table 2: Quantitative comparison results for ablation study on the proposed losses.
Figure 5: Comparison of ORM results across all methods.
Appearance metrics
Geometry metrics
Time
Angles
PSNR ↑
SSIM ↑
LPIPS ↓
Clip-Sim ↑
CD ↓
F-Score ↑
seconds ↓
CVCR (3°)
21.8628
0.8971
0.1455
0.7764
0.02030
0.4203
208.1
CVCR (5°)
21.8407
0.8966
0.1453
0.7753
0.02023
0.4212
163.3
CVCR (10°)
21.3687
0.8879
0.1527
0.7591
0.02134
0.4118
129.5
CVCR (15°)
20.8415
0.8763
0.1671
0.7459
0.02221
0.3936
110.8
Table 3: Ablation study of angular intervals between adjacent views in the CVCR module.
Figure 6: After applying the CVCR module to Unique3D.
Diffusion priors have recently demonstrated strong capability in enhancing the quality of sparse-view 3D reconstruction by augmenting training views at novel viewpoints, but they inevitably introduce hallucinated content -- artifacts inconsistent with the input views -- into the final 3D model. To address this challenge, we propose Hallucination-Aware Diffusion prior (HAD), which estimates pixel-wise hallucination score maps for augmented images by leveraging multi-view reasoning capabilities from a feedforward novel view synthesis (NVS) network pre-trained on large-scale 3D data. These hallucination scores enable selective masking of unreliable pixels during the progressive 3D reconstruction procedure, preventing the introduction of non-existent artifacts into the 3D model. To further enhance performance, we create multiple versions of augmented images at each novel view by conditioning the diffusion prior on different input views, which are then fused into a final image that leverages the broader context across all input views. We show that our method substantially reduces hallucination artifacts in diffusion-assisted 3D reconstruction, thereby achieving state-of-the-art performance across multiple benchmarks on novel view synthesis. Our project are publicly available at project website.
Reconstructing complete, high-quality meshes of large-scale scenes from unposed multi-view images requires faithful geometric grounding while plausibly completing surfaces that are unobserved or only sparsely captured. We present WorldRecon, a generative multi-view reconstruction framework for recovering complete, large-scale 3D scene meshes from unposed RGB images. WorldRecon uses a coarse, noisy global point cloud and camera poses estimated from the input images by a reconstruction model as a global scaffold for whole-scene generation and partitions it into overlapping spatial blocks. To robustly leverage these imperfect geometric priors, we equip the native 3D generative model TRELLIS.2 with a dedicated reconstruction-based conditioning architecture that generates each block from its local point-cloud crop and block-relevant image observations, producing sparse structure, detailed geometry, and PBR materials. To maintain scene-level coherence, we introduce neighbour-aware latent inpainting and a causal world-generation procedure that propagates structural, geometric, and material context across overlapping blocks. Together, these designs scale multi-view generative reconstruction to large-scale scenes, yielding coherent, unified meshes. Extensive experiments demonstrate that WorldRecon achieves state-of-the-art reconstruction performance, reducing Chamfer distance by 14.9% and 43.1% while improving F-score by 17.3% and 12.0% over the strongest baselines on synthetic and real-world indoor scenes, respectively, thereby demonstrating faithful and complete generative reconstruction as well as strong sim-to-real generalization. Project page: https://kszpxxzmc.github.io/WorldRecon-website/.
Mengchen Zhang, Yang Li, Chunshi Wang +2
Zhejiang University · Tencent Hunyuan · Fudan University
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/
Siang-Ling Zhang, Huai-Hsun Cheng, Tsung-Ju Yang +1