3D-Fixer: In-Place 3D Completion

• 3D-Fixer is a set of 3D repair and completion pipelines that restore incomplete scenes using fragmented geometry as a spatial anchor.
• It employs a dual-stage diffusion framework with a Coarse Structure Completer and Fine Shape Refiner to resolve occlusion ambiguity and enhance details.
• The method achieves superior performance on benchmarks, lowering scene Chamfer Distance and improving F-Score compared to other state-of-the-art approaches.

3D-Fixer designates a class of 3D repair and completion pipelines that retain an existing spatial carrier while correcting degradation or inferring missing structure. In its explicit 2026 formulation, 3D-Fixer is a novel in-place completion paradigm for 3D scenes from a single image: it extends 3D object generative priors to generate complete 3D assets conditioned on the partially visible point cloud at the original locations, using fragmented geometry as a spatial anchor rather than explicit pose alignment (Yin et al., 6 Apr 2026). In related usages, a “3D-Fixer” pipeline also denotes an ArtiFixer-style system for enhancing sparse-view 3D reconstructions through opacity-aware generative modeling and auto-regressive distillation, and a FoR$^2$M-derived module for the recognition and repair of foldings in mesh surfaces (Lutio et al., 28 Feb 2026, Sfikas et al., 2022).

1. Scope and problem settings

A common source of confusion is that “3D-Fixer” names both a specific single-image scene-completion method and a broader repair-oriented tool concept. In the cited works, the term spans three problem settings: compositional 3D scene generation from a single view, enhancement and extension of sparse-view 3D reconstruction, and repair of foldings in triangular meshes (Yin et al., 6 Apr 2026, Lutio et al., 28 Feb 2026, Sfikas et al., 2022).

Usage	Input	Output or target
3D-Fixer	A single color image $I$, instance masks $\{M_i\}_{i=1}^N$, and fragmented point clouds $G^{(i)}_{\mathrm{frag}}$	A complete 3D scene composed of $N$ dense point clouds (or meshes) $G^{(i)}_{\mathrm{full}}$, aligned in the camera coordinate frame
ArtiFixer-style “3D-Fixer”	Sparse reference images, an optional text prompt, a degraded novel-view rendering $\mathbf{I}_{deg}$, per-pixel opacities $\mathbf{O}(\mathbf{p})$, and camera ray embeddings $\mathbf{R}$	Enhanced frames, arbitrarily long novel-view sequences, or pseudo-supervision for refining 3D Gaussians
FoR$^2$M as a 3D-Fixer module	A triangular mesh $I$0	Recognition and repair of mesh surface foldings, removal of folded regions, and hole filling

For the single-image method, existing approaches mainly fall into two categories: feed-forward generation methods and per-instance generation methods. The former directly predict 3D assets with explicit 6DoF poses through efficient network inference, but they generalize poorly to complex scenes. The latter improve generalization through a divide-and-conquer strategy, but suffer from time-consuming pose optimization. 3D-Fixer is introduced to bridge this gap through in-place completion.

2. In-place completion from a single image

The single-image 3D-Fixer formulation takes as input a single color image $I$1, a set of instance masks $I$2, and for each instance a partially observed fragmented point cloud

$I$3

obtained by mono-depth or point-cloud estimators such as MoGe-2 and VGGT. The output is a complete 3D scene composed of $I$4 dense point clouds or meshes,

$I$5

aligned in the camera coordinate frame. The latent representation in the diffusion model is denoted by $I$6, with $I$7 indexing the diffusion time step $I$8; the partial geometry is written as $I$9, and the full target geometry as $\{M_i\}_{i=1}^N$0 (Yin et al., 6 Apr 2026).

For each object instance, 3D-Fixer learns a conditional diffusion model

$\{M_i\}_{i=1}^N$1

where $\{M_i\}_{i=1}^N$2. In the score-based view, the one-step reverse is

$\{M_i\}_{i=1}^N$3

with $\{M_i\}_{i=1}^N$4 a U-Net–style transformer that predicts the noise.

The central mechanism is a coarse-to-fine generation scheme to resolve boundary ambiguity under occlusion. The method trains two diffusion models jointly, a Coarse Structure Completer and a Fine Shape Refiner, under the combined objective

$\{M_i\}_{i=1}^N$5

where $\{M_i\}_{i=1}^N$6. Here $\{M_i\}_{i=1}^N$7 is the Flow-Matching denoising loss and $\{M_i\}_{i=1}^N$8 is the Occlusion-Robust Feature Alignment loss. The use of fragmented geometry as a spatial anchor preserves layout fidelity without expensive pose optimization.

3. Conditioning architecture, ORFA, and ARSG-110K

At the architectural core of 3D-Fixer is a dual-branch conditioning network. The “generative branch” is the original TRELLIS DiT encoder–decoder and is frozen. The “context branch” is trainable and ingests two inputs: the voxelized partial geometry $\{M_i\}_{i=1}^N$9 in a $G^{(i)}_{\mathrm{frag}}$0 grid compressed to a $G^{(i)}_{\mathrm{frag}}$1 latent by a 3D VAE, and global image features via DINOv2 and MoGe-2 tokens. At each transformer layer, geometry features attend to generative latents via cross-attention; reciprocally, generative features inject through residual addition. The Occlusion-Robust Feature Alignment module aligns context-branch activations $G^{(i)}_{\mathrm{frag}}$2 to teacher activations $G^{(i)}_{\mathrm{frag}}$3 through cosine similarity, thereby stabilizing diffusion under heavy occlusion (Yin et al., 6 Apr 2026).

The basic computation unit is a standard DiT block with hidden dimension $G^{(i)}_{\mathrm{frag}}$4, 8 attention heads, and MLP inner-dim $G^{(i)}_{\mathrm{frag}}$5. Both coarse- and fine-stage models use 12 such layers. Training further uses mixed geometry noise: with probability $G^{(i)}_{\mathrm{frag}}$6, estimated depth is mixed with ground-truth depth. Sampling uses classifier-free guidance $G^{(i)}_{\mathrm{frag}}$7 and 25 diffusion steps at inference.

The data regime is defined by ARSG-110K, described as the largest scene-level dataset to date. It contains 110K procedurally generated scenes rendered with Blender Cycles, 180K+ high-quality 3D models from Objaverse, 3D-Future, HABITAT-HSSD, and ABO, and 30 random cameras per scene for a total of 3M RGB images. The annotations include camera intrinsics and extrinsics, per-pixel instance masks, and full 3D meshes with 6DoF poses. Diversity is specified as 5–20 objects per scene, 1,000 HDR maps, 5,000 textures, and inter-object occlusion in at least 40% of masks. The training schedule comprises 100K steps of pre-training on object-level data, followed by 80K steps each for the Coarse Completer and Fine Refiner and 90K steps for the 3D Texturer, all with batch size 128, AdamW, and learning rate $G^{(i)}_{\mathrm{frag}}$8.

4. Quantitative behavior, ablations, and metric interpretation

On the MIDI Testset, the reported quantitative metrics are: Scene Chamfer Distance $G^{(i)}_{\mathrm{frag}}$9 for 3D-Fixer versus $N$0 for MIDI and $N$1 for Gen3DSR; Scene F-Score @0.1 of $N$2 for 3D-Fixer versus $N$3 for MIDI and $N$4 for Gen3DSR; IoU of $N$5 for 3D-Fixer versus $N$6 for MIDI and $N$7 for Gen3DSR; and inference time of $N$8 for 3D-Fixer versus $N$9 for MIDI and $G^{(i)}_{\mathrm{full}}$0 for Gen3DSR. On real ScanNet frames, the reported Scene Chamfer Distance is $G^{(i)}_{\mathrm{full}}$1 for 3D-Fixer, compared with $G^{(i)}_{\mathrm{full}}$2 for MIDI and $G^{(i)}_{\mathrm{full}}$3 for Gen3DSR. Under heavy occlusion, 3D-Fixer is reported to recover complete chair and table geometries where both MIDI and Gen3DSR fail or hallucinate misaligned shapes (Yin et al., 6 Apr 2026).

The ablations isolate the importance of the method’s principal components. Removing Coarse-to-Fine raises $G^{(i)}_{\mathrm{full}}$4 from $G^{(i)}_{\mathrm{full}}$5. Removing ORFA alignment raises training instability. Reducing layers to 6 degrades $G^{(i)}_{\mathrm{full}}$6 by approximately $G^{(i)}_{\mathrm{full}}$7. The paper’s own explanation is threefold: in-place completion uses the fragmented point cloud as a spatial anchor and preserves layout without expensive pose optimization; Coarse-to-Fine decouples boundary estimation from detail generation, resolving occlusion ambiguity; and ORFA aligns occluded-scene training with a clean-scene teacher.

A useful corrective to a simplistic reading of the benchmark is that the reported superiority is not uniform across every metric. 3D-Fixer improves Scene Chamfer Distance and Scene F-Score @0.1, but the reported IoU is $G^{(i)}_{\mathrm{full}}$8 for 3D-Fixer and $G^{(i)}_{\mathrm{full}}$9 for MIDI. This suggests that the claim of state-of-the-art geometric accuracy is supported primarily by the distance- and coverage-oriented measures reported in the paper rather than by strict dominance on every evaluation axis.

5. ArtiFixer-style 3D-Fixer for sparse-view reconstruction enhancement

In a distinct formulation, a “3D-Fixer” pipeline can be organized as a two-stage system for enhancing and extending 3D reconstruction with auto-regressive diffusion models. Stage 1 is a bidirectional teacher that takes sparse reference images, an optional text prompt, a degraded novel-view rendering $\mathbf{I}_{deg}$0 produced by a fast 3D engine such as 3D Gaussian Splatting, per-pixel opacities $\mathbf{I}_{deg}$1, and camera ray embeddings $\mathbf{I}_{deg}$2. Its goal is to “clean up” $\mathbf{I}_{deg}$3 and inpaint entirely unseen regions. Stage 2 is a causal auto-regressive student obtained by adding a causal attention mask to the finetuned bidirectional model and distilling it via one-step distribution-matching into a 4-step latent diffusion model. At inference, the student generates arbitrarily long sequences frame by frame, conditioning on previous frames via key/value caching (Lutio et al., 28 Feb 2026).

The defining technical device is the opacity-mixing strategy. Rather than starting from pure Gaussian noise or simple channel-concatenation conditioning, the method forms

$\mathbf{I}_{deg}$4

where $\mathbf{I}_{deg}$5 is the VAE latent of the degraded rendering, $\mathbf{I}_{deg}$6 is the max-pooled opacity map in latent resolution, and $\mathbf{I}_{deg}$7. In highly opaque regions, generation starts from the true latent; in zero-opacity regions, it starts from noise. The teacher is built on Wan 2.1 T2V-14B and uses latent video tokens, degraded mixed latents, reference-view tokens, Plücker-raymaps, opacity embeddings, and optional text conditioning. Training uses conditional flow-matching; student training adds Diffusion-Forcing perturbation, Self Forcing–style rollout, and one-step Distribution Matching Distillation.

The generated views can serve as pseudo-supervision for refining 3D Gaussians $\mathbf{I}_{deg}$8 through standard 3DGS gradient descent with AdamW for a few hundred iterations; optionally, the video model may be re-applied for a final pass (“ArtiFixer3D+”). On commonly benchmarked datasets, the reported PSNR gains are $\mathbf{I}_{deg}$9 over Difix3D+ on Nerfbusters and DL3DV, $\mathbf{O}(\mathbf{p})$0 versus $\mathbf{O}(\mathbf{p})$1 on Mip-NeRF360 3-view splits, and $\mathbf{O}(\mathbf{p})$2 versus $\mathbf{O}(\mathbf{p})$3 on novel-content DL3DV sparse against GenFusion. The implementation is built on $\mathbf{O}(\mathbf{p})$4 GPUs; the teacher uses 15K iterations at $\mathbf{O}(\mathbf{p})$5, and the student uses 5K initialization iterations followed by 2K rollout plus DMD at $\mathbf{O}(\mathbf{p})$6. The stated limitations are inference speed relative to native 3DGS realtime splatting, block-chunk latency, semantically plausible but incorrect geometry in extremely under-observed scenes, and a speed–fidelity trade-off in fully novel regions.

6. FoR$\mathbf{O}(\mathbf{p})$7M as a mesh-surface 3D-Fixer module

A third usage treats 3D-Fixer as a mesh-repair tool into which FoR$\mathbf{O}(\mathbf{p})$8M’s fold-detection and repair pipeline can be embedded. The underlying representation is a triangular mesh $\mathbf{O}(\mathbf{p})$9, where $\mathbf{R}$0 is the vertex set and $\mathbf{R}$1 is the set of oriented triangular faces. A folded region consists of two subsets of faces: $\mathbf{R}$2, the intersecting faces, and $\mathbf{R}$3, the inward-oriented faces. A face $\mathbf{R}$4 is in $\mathbf{R}$5 if there exists $\mathbf{R}$6, $\mathbf{R}$7, such that $\mathbf{R}$8 and neither $\mathbf{R}$9 nor $^2$0 share only a common edge or vertex. A face is inward-oriented when the parity of intersections between the ray $^2$1, $^2$2, and the mesh is odd. A connected component is classified as folded if a sufficiently large fraction of its faces satisfy the odd-parity test (Sfikas et al., 2022).

Detection proceeds in two phases: an axis-aligned bounding-box filter, then an exact triangle–triangle test via Möller–Trumbore. AABB overlap quickly rejects non-candidates, while the exact test handles only true geometric intersections. Worst-case complexity is $^2$3 triangle–triangle AABB tests, with fewer exact tests; in practice, spatial hashing reduces this to $^2$4 where $^2$5. After computing $^2$6, the method removes all intersecting faces and any protruding face with at least $^2$7 of its neighbors in $^2$8. The remaining faces are partitioned into connected components via adjacency lists and flood-fill. For each component $^2$9, if more than $I$00 of faces have odd parity, with $I$01, the component is marked folded and removed.

Gap reconstruction has two stages. First, removed triangles are split along stored intersection line segments $I$02, and the subtriangles on the kept side are stitched to the boundary of the kept mesh. Second, each remaining hole boundary $I$03 is filled by solving a discrete Laplace minimal-surface patch problem: $I$04 or, equivalently,

$I$05

The patch is $I$06-continuous with the existing mesh, and using a half-edge structure ensures each new triangle is manifold.

The implementation notes specify support for OBJ and PLY, adjacency via half-edge or winged-edge, and spatial hash or uniform grid acceleration. The reported benchmarks are: $I$07 total for a small model of approximately 1K vertices and 2K faces; approximately $I$08 for a medium model of approximately 50K vertices and 100K faces; an intersection module alone that is $I$09 faster than pure Möller–Trumbore on planar meshes; and a success rate of $I$10 on the tested public meshes, including Stanford bunny and dragon, Purdue ESB, Princeton PSB, McGill, and SHREC. The stated limitations are inability to reconstruct topology completely lost, possible oversmoothing of high-frequency detail by the Laplacian patch, and worst-case $I$11 time if the spatial index degrades.