Wonder3D++: 3D Mesh Reconstruction from One Image

• The paper introduces a novel method that jointly models multi-view normal maps and color images via cross-domain diffusion and cascaded explicit-mesh reconstruction.
• It leverages multi-view cross-domain attention to align geometry and photorealistic textures without relying on explicit reprojection losses.
• The cascaded pipeline—comprising geometric initialization, coarse optimization, and iterative high-resolution refinement—achieves competitive Chamfer and IoU metrics on reconstruction benchmarks.

Wonder3D++ is a single-image 3D reconstruction method for generating high-fidelity textured meshes from a single-view image through cross-domain diffusion and a cascaded explicit-mesh reconstruction pipeline. It directly models the joint distribution of multi-view normal maps and corresponding color images, uses multi-view cross-domain attention to exchange information across views and modalities, and reconstructs an explicit mesh through a coarse-to-fine procedure that combines differentiable rendering, inconsistency-aware optimization, remeshing, UV-based texture generation, and iterative high-resolution refinement (Yang et al., 3 Nov 2025). In the Wonder3D lineage, it represents a development beyond the original Wonder3D, whose formulation centered on cross-domain diffusion plus geometry-aware SDF optimization rather than the later cascaded explicit-mesh extraction pipeline (Long et al., 2023).

1. Definition and problem setting

Wonder3D++ addresses single-image, high-fidelity textured mesh reconstruction. The stated objective is to infer both visible and occluded geometry with high geometric detail and photorealistic texture while remaining efficient and view-consistent (Yang et al., 3 Nov 2025). The method is positioned between two prior tendencies in the literature. On one side are SDS-based per-shape optimization pipelines, which inherit 2D diffusion priors but require tens of minutes to hours of optimization and often suffer Janus or multi-face artifacts and inconsistent geometry. On the other side are direct feed-forward inference models trained natively in 3D, which are fast but often yield low-detail geometry or textures and may misalign with the input image (Yang et al., 3 Nov 2025).

The central formulation treats geometry and appearance jointly. Let $N^{1:K}$ denote $K$-view normal maps and $X^{1:K}$ the corresponding $K$-view color images. Wonder3D++ models the 3D asset distribution through multi-view 2D observations as

$p_a(z) = p_{nc}(N^{1:K}, X^{1:K} \mid y),$

where $y$ is the input image. The reverse process is defined as a joint Markov chain over $(N, X)$, initialized by Gaussian noise (Yang et al., 3 Nov 2025).

This design directly encodes a specific stance on single-view reconstruction: rather than optimizing a 3D representation against a 2D prior at test time, the model learns a reusable multi-view, multi-domain generative prior once and then reconstructs geometry from the generated normal and color observations. A plausible implication is that the method treats geometric supervision as a first-class generative target instead of a downstream by-product of RGB synthesis.

2. Cross-domain diffusion architecture

The generator is built on a Stable Diffusion-style latent U-Net with transformer blocks. Conditioning is provided through four components: the single input image, the camera parameters of $K=6$ target views, a domain switcher, and a camera-type switcher (Yang et al., 3 Nov 2025).

The input image enters through two routes. A CLIP-Vision embedding is injected through cross-attention, and a VAE-encoded latent is concatenated to the noise latent in an image-to-image-variations configuration. The camera parameters correspond to six target views covering front, back, left, right, front-left, and front-right around the input-view-aligned frame, using fixed-radius azimuth sampling in the $X_v$–$Y_v$ plane. The domain switcher $K$0 selects the output domain, while the camera-type switcher $K$1 controls projection type. Each switcher is positional-encoded and fused with the time embedding (Yang et al., 3 Nov 2025).

The reverse diffusion objective follows the latent-diffusion noise-prediction form. If $K$2 is a clean latent and $K$3 is the forward process, the reverse model predicts

$K$4

with conditioning

$K$5

The training loss is

$K$6

Wonder3D++ fine-tunes a multi-view variant of Stable Diffusion in three stages to reach the joint cross-domain, multi-view target distribution (Yang et al., 3 Nov 2025).

The backbone is shared across the normal and color domains. The domain switcher selects which latent is produced, but the feature space and denoising machinery are coupled. This shared-latent arrangement is important for interpreting Wonder3D++: geometry and appearance are not modeled by isolated branches but by a single denoising system with explicit domain control.

3. Multi-view cross-domain attention and consistency mechanisms

Consistency is enforced primarily through attention design rather than explicit reprojection losses during diffusion training. The model uses cross-domain attention and global multi-view attention, with camera conditioning supplied per view (Yang et al., 3 Nov 2025).

For cross-domain attention, let $K$7 and $K$8 be latent features for the normal and color domains within a shared transformer block. The module concatenates domain features in the key and value paths:

$K$9

The updates are

$X^{1:K}$0

This exchanges evidence between geometry and appearance at every denoising layer (Yang et al., 3 Nov 2025).

For multi-view attention, tokens from all $X^{1:K}$1 views are stacked with view-specific positional encodings and camera metadata, and self-attention becomes global across views. The data describe a conceptual form

$X^{1:K}$2

where $X^{1:K}$3 can encode relative view indices or camera poses and $X^{1:K}$4 distinguishes normal versus color tokens; however, the implementation uses view concatenation and conditioning on camera parameters rather than requiring explicit bias matrices (Yang et al., 3 Nov 2025).

An important negative statement in the formulation is that the diffusion model is trained via denoising loss and does not use explicit cross-view reprojection or photometric consistency terms during diffusion training. Consistency is enforced implicitly by multi-view attention and explicit conditioning on camera poses, while explicit multi-view normal and mask supervision appear later during mesh extraction (Yang et al., 3 Nov 2025). This distinguishes Wonder3D++ from pipelines in which geometric consistency is imposed directly in the generative objective.

Cross-domain attention also clarifies the relation to the earlier Wonder3D system. The original Wonder3D already used a cross-domain diffusion model that generated multi-view normal maps and corresponding color images and introduced multi-view cross-domain attention for consistency (Long et al., 2023). Wonder3D++ preserves that core idea but redesigns the joint Normal–RGB sampling regime, augments conditioning with a camera-type switcher, and situates the generative stage within a more elaborate reconstruction cascade (Yang et al., 3 Nov 2025).

4. Cascaded coarse-to-fine mesh extraction

Wonder3D++ reconstructs an explicit mesh from $X^{1:K}$5 at $X^{1:K}$6 through a three-stage cascade: geometric initialization, inconsistency-aware coarse reconstruction, and iterative high-resolution refinement (Yang et al., 3 Nov 2025).

The geometric initialization stage uses two monocular normal maps, front and back, integrates them to depth, back-projects them to point clouds with normals, and fuses them by Poisson surface reconstruction to obtain an initial mesh. A concavity check is used as a safeguard: if the mean depth over a central region and the full map indicates likely concavity, the initializer falls back to a sphere to avoid topologically inconsistent Poisson fusion (Yang et al., 3 Nov 2025).

The coarse reconstruction stage optimizes the initialized mesh $X^{1:K}$7 using differentiable rendering with nvdiffrast. Normals $X^{1:K}$8 and masks $X^{1:K}$9 are rendered, and the geometry is refined by minimizing

$K$0

where

$K$1

The geometry-aware normal term is

$K$2

with normalized weights

$K$3

Here $K$4 is the current mesh normal, $K$5 the generated normal seen from view $K$6, $K$7 the viewing ray, $K$8 a visibility mask, and $K$9 a small negative threshold. Continuous remeshing follows each optimization iteration through edge split, collapse, flip, and smoothing (Yang et al., 3 Nov 2025).

Texture construction is performed in UV space. First, multi-view colors are fused onto vertices by weighted averaging,

$p_a(z) = p_{nc}(N^{1:K}, X^{1:K} \mid y),$0

and invisible vertices are colored by propagation along the mesh graph. Second, the mesh is unwrapped to a UV atlas, multi-view images are reprojected to UV, and the results are blended with the initial UV texture, with dilation used to fill small gaps (Yang et al., 3 Nov 2025).

The final stage is iterative high-resolution refinement. The coarse mesh is rendered from four views, and those rendered pairs, together with the low-resolution generated $p_a(z) = p_{nc}(N^{1:K}, X^{1:K} \mid y),$1, are fed into a cross-domain multi-view enhancement module. This module is described as a ControlNet conditioned SD branch with cross-domain attention and an IP-Adapter for the input image; it uses DDIM inversion for the rendered pairs and produces enhanced normals and colors at $p_a(z) = p_{nc}(N^{1:K}, X^{1:K} \mid y),$2. The mesh is then refined again against the enhanced pairs using the same differentiable rendering objective, in a loop that progressively sharpens geometry and texture (Yang et al., 3 Nov 2025).

A compact interpretation of the mesh objective is

$p_a(z) = p_{nc}(N^{1:K}, X^{1:K} \mid y),$3

The Poisson initializer is likewise expressed as minimizing

$p_a(z) = p_{nc}(N^{1:K}, X^{1:K} \mid y),$4

with $p_a(z) = p_{nc}(N^{1:K}, X^{1:K} \mid y),$5, before extracting an isosurface (Yang et al., 3 Nov 2025).

5. Training curriculum, data, and inference characteristics

Wonder3D++ is fine-tuned on a cleaned single-object subset of Objaverse comprising approximately $p_a(z) = p_{nc}(N^{1:K}, X^{1:K} \mid y),$6 objects. BlenderProc renders six azimuthal views per object under both orthographic and perspective cameras: front, back, left, right, front-left, and front-right. Inputs are normalized to unit scale and center, and multi-domain pretraining additionally uses masks as a weak shape cue (Yang et al., 3 Nov 2025).

The training schedule has three stages. In Stage 1, multi-domain pretraining removes the domain switcher and cross-domain attention, converts self-attention to multi-view, and trains prediction of multi-view targets from a single view within the same domain, namely colors, normals, or masks. In Stage 2, mixed-domain fine-tuning adds the domain switcher and trains prediction of either multi-view colors or multi-view normals from a single-view color image. In Stage 3, cross-domain alignment adds cross-domain attention, freezes earlier weights, and jointly predicts both color and normal multi-view outputs conditioned on a single-view color image (Yang et al., 3 Nov 2025).

The main diffusion training uses $p_a(z) = p_{nc}(N^{1:K}, X^{1:K} \mid y),$7 resolution, total batch size $p_a(z) = p_{nc}(N^{1:K}, X^{1:K} \mid y),$8, and $p_a(z) = p_{nc}(N^{1:K}, X^{1:K} \mid y),$9 steps, requiring approximately $y$0 days on $y$1 Huawei Kunpeng 910B GPUs. The cross-domain multi-view enhancement module is trained at $y$2 for $y$3 steps with batch size $y$4, taking approximately $y$5 days on $y$6 910B GPUs. The base model starts from Stable Diffusion Image Variations with image conditioning retained (Yang et al., 3 Nov 2025).

At inference time, the pipeline proceeds by encoding the input with CLIP-Vision and VAE, selecting the camera type $y$7, preparing the six camera poses $y$8, sampling multi-view normals and colors by toggling the domain switcher, initializing the mesh from normals, optimizing the coarse geometry with remeshing, generating the initial UV texture, and then running one or more iterative refinement rounds (Yang et al., 3 Nov 2025).

The reported runtime on a single A100 is approximately $y$9 seconds end-to-end, decomposed into $(N, X)$0 seconds for multi-view diffusion, $(N, X)$1 seconds for initialization, $(N, X)$2 seconds for coarse reconstruction, and $(N, X)$3 seconds for iterative refinement. The account attributes the practical speed-up to the explicit mesh plus remeshing design, with complexity scaling with vertex count and image resolution squared rather than cubed (Yang et al., 3 Nov 2025).

The following summary condenses the pipeline components and their stated functions.

Component	Stated mechanism	Stated role
Multi-view diffusion	Shared latent U-Net with domain and camera-type switchers	Generate $(N, X)$4 normals and colors
Cross-domain attention	Shared $(N, X)$5 banks over normal and color features	Align geometry and appearance
Geometric initialization	Normal integration, back-projection, Poisson fusion, concavity fallback	Produce initial explicit mesh
Coarse optimization	nvdiffrast, $(N, X)$6	Refine mesh geometry
UV-based texture generation	Vertex fusion, graph propagation, UV reprojection and blending	Build coherent texture
High-resolution refinement	ControlNet-style enhancement, IP-Adapter, DDIM inversion	Improve $(N, X)$7 geometry and texture

6. Evaluation, relation to Wonder3D, and limitations

On Google Scanned Objects, Wonder3D++ reports the lowest Chamfer distance and highest IoU among the listed single-view reconstruction baselines. The orthographic setting achieves CD $(N, X)$8 and IoU $(N, X)$9, while the perspective setting achieves CD $K=6$0 and IoU $K=6$1. The comparison list includes RealFusion, Magic123, One-2-3-45, Point-E, Shap-E, Zero123, SyncDreamer, OpenLRM, InstantMesh, CRM, Unique3D, Era3D, and the original Wonder3D (Yang et al., 3 Nov 2025).

For novel view synthesis on the same benchmark, Wonder3D++ reports PSNR $K=6$2, SSIM $K=6$3, and LPIPS $K=6$4 in the orthographic setting, and PSNR $K=6$5, SSIM $K=6$6, and LPIPS $K=6$7 in the perspective setting. These values are reported as improvements over SyncDreamer, SV3D, Unique3D, Era3D, and Wonder3D (Yang et al., 3 Nov 2025).

The paper attributes several ablation findings to specific components. Cross-domain attention increases alignment between normals and colors and improves geometry fidelity. The geometry-aware normal loss sharpens surfaces and reduces artifacts. The camera-type switcher prevents distortions when the input projection differs from training defaults. The multi-stage training schedule stabilizes convergence and improves multi-view consistency. The cascaded refinement materially improves both geometry and texture over single-pass extraction and over the implicit SDF reconstruction used in Wonder3D (Yang et al., 3 Nov 2025).

The relationship to the original Wonder3D is structurally explicit. Wonder3D introduced a cross-domain diffusion model that generated multi-view normal maps and corresponding color images, enforced consistency with multi-view cross-domain attention, and recovered surfaces through geometry-aware normal fusion with an implicit SDF optimized from normals, colors, and masks (Long et al., 2023). Wonder3D++ modifies all three of those axes. It adds a camera-type switcher and a multi-domain pretraining curriculum, redesigns cross-domain attention for joint Normal–RGB sampling, replaces implicit SDF and NeUS-based optimization with explicit mesh optimization plus continuous remeshing and a cascaded refinement loop, and adds a cross-domain multi-view enhancement module based on ControlNet, IP-Adapter, and DDIM inversion (Yang et al., 3 Nov 2025). This suggests that the “++” designation is not merely nominal; it denotes a change in reconstruction substrate from implicit field optimization to explicit coarse-to-fine mesh optimization.

The limitations are also stated directly. View extrapolation under extreme occlusion and highly complex, concave topologies can still fail, especially when six views are insufficient to disambiguate geometry. Thin structures and severe self-occlusion remain challenging. Materials with strong specularities, translucency, or transparency are not explicitly modeled, and normal supervision together with Lambertian assumptions can cause texture bleeding or incorrect normal cues. Performance also depends on correct camera-type selection, since mis-specified projection causes distortions even though the switcher mitigates the issue (Yang et al., 3 Nov 2025).

Implementation resources are available through the Wonder3D_Plus repository, and the recommended configuration follows the three-stage curriculum, six-view orthographic and perspective rendering, $K=6$8 base-model training, $K=6$9 enhancement training, concavity-aware initialization, remeshing during coarse optimization, and one to two refinement rounds for best detail (Yang et al., 3 Nov 2025).