ARSG-110K: Large-Scale Synthetic 3D Scene Dataset

• ARSG-110K is a synthetic, asset-rich 3D scene dataset comprising 110K scenes and over 3M rendered views with precise object meshes and 6DoF poses.
• It enables compositional scene generation and in-place completion by providing detailed annotations including RGB images, depth maps, instance masks, and camera parameters.
• The dataset is procedurally generated in Blender, yielding diverse occlusion patterns and complex scene layouts that challenge and enhance 3D model training.

Searching arXiv for the primary paper and a few related scene-generation references so the article can cite them precisely. ARSG-110K is a large-scale synthetic 3D scene dataset introduced alongside 3D-Fixer for compositional 3D scene generation from a single image. In the source paper, the name is expanded as Asset-Rich Scene Generation–110K, and the dataset is positioned as a training resource and benchmark built to address a data bottleneck in scene-level 3D generation: existing datasets are described as being either too small in the number of scenes, lacking object-level 3D assets with accurate poses, or not sufficiently aligned and high-fidelity at the object level (Yin et al., 6 Apr 2026). ARSG-110K comprises over 110K diverse scenes and 3M annotated images with high-fidelity 3D ground truth, with complete 3D mesh models for every asset, accompanied by their precise translation and rotation matrices in the scene coordinate (Yin et al., 6 Apr 2026).

1. Origin and stated purpose

ARSG-110K was created to support compositional 3D scene generation and, more specifically, the in-place completion paradigm introduced by 3D-Fixer. The motivating problem is that single-view scene generation requires simultaneous recovery of scene layout and 3D assets, while prior approaches either directly predict 3D assets with explicit 6DoF poses but generalize poorly to complex scenes, or improve generalization through per-instance processing but incur time-consuming pose optimization (Yin et al., 6 Apr 2026). ARSG-110K is presented as the dataset infrastructure intended to bridge this gap.

The dataset is described as the “missing” large-scale, asset-rich, scene-level resource for training and benchmarking compositional 3D scene generation and in-place completion from a single image. Its design goals are explicit: combine large scale, high-quality watertight object meshes, precise 6DoF poses and camera parameters, and complex occlusion patterns and varied layouts (Yin et al., 6 Apr 2026). This suggests that ARSG-110K is not merely a larger scene corpus, but a dataset optimized for methods that must infer complete geometry from fragmented observations while preserving scene layout.

Within the 3D-Fixer framework, ARSG-110K functions both as supervision and as a source of synthetic scene complexity. The dataset provides occluded scene views, object-level geometry, camera parameters, depth, and instance masks, enabling the model to learn completion of partial geometry at original locations rather than through explicit pose alignment (Yin et al., 6 Apr 2026).

2. Scale, composition, and coverage

The paper states that ARSG-110K contains over 110,000 unique scenes, over 3,000,000 rendered views, and over 180,000 unique 3D object assets, with 30 random camera views per scene (Yin et al., 6 Apr 2026). Each scene is densely populated with 5–20 individual assets, purposefully creating complex object-to-object occlusion scenarios. This density is central to the dataset’s intended difficulty.

ARSG-110K draws assets from several repositories. The paper lists Objaverse++ / Objaverse, 3D-FUTURE, HSSD, and ABO as source assets, and reports the following aggregate scale (Yin et al., 6 Apr 2026):

Component	Reported value
Objects	180K+
Scenes	110K
Images	3M

The paper also compares ARSG-110K with prior scene-level datasets and asset collections, emphasizing scale and asset richness (Yin et al., 6 Apr 2026):

Dataset	Objects / Scenes / Images
Scan2CAD	14.2K / 706 / 2.5M
OpenRooms	16.0K / 706 / 2.5M
MetaScenes	15.4K / 706 / 2.5M
R3DS	19.1K / 370 / 194K
CAD-Estate	100K / 19.5K / –
BVS	6.7K / 1000 / –
HSSD-200	18.7K / 211 / –
3D-FRONT	13.2K / 18.9K / 20K
ARSG-110K	180K+ / 110K / 3M

The paper’s interpretation of this comparison is that ARSG-110K has more scenes than the listed alternatives and more unique object models, while also providing exact meshes and poses by construction (Yin et al., 6 Apr 2026). A plausible implication is that the dataset is intended to support scale-sensitive training regimes that were previously difficult to realize with smaller asset-rich datasets.

3. Scene construction and procedural generation

ARSG-110K is fully synthetic and is constructed via automated procedural scene generation followed by photorealistic rendering in Blender Cycles (Yin et al., 6 Apr 2026). The pipeline begins by creating a floor plane and then probabilistically placing 0 to 4 additional planes around it as walls to simulate both indoor and outdoor environments. Floor and walls are assigned random material textures from a pool of 5K+ textures, and lighting is sampled from over 1K HDR maps from BlenderKit (Yin et al., 6 Apr 2026).

The asset placement process is also explicitly procedural. The dataset generation script randomly samples 20 3D assets from the 180K asset pool. For each asset, the mesh is normalized, a random rotation around the z-axis is applied, and a random uniform scaling factor $s \in [0.5, 2.0]$ is assigned. Objects are then placed sequentially with collision avoidance. For each candidate object, up to 100 placement attempts are made; if no collision-free placement is found, the object is dropped. As a result, each scene ends up with 5–20 placed assets (Yin et al., 6 Apr 2026).

The generation process can be summarized as follows:

1. Create a floor plane.
2. Probabilistically add 0 to 4 wall planes.
3. Randomly assign material textures.
4. Randomly select an HDR map for environmental lighting.
5. Sample 20 assets from the asset pool.
6. Normalize, rotate around the z-axis, and scale each asset.
7. Place assets sequentially with collision checks and up to 100 attempts each.
8. Store scene-level ground truth, including asset identity and final transform (Yin et al., 6 Apr 2026).

The source paper states that this design yields varied inter-object relationships and partial occlusions. On the test set constructed with the same procedure, 40.18% of instance masks have more than one 8-connected component, and 10.89% have more than four (Yin et al., 6 Apr 2026). This is a direct quantitative indication that visible object support is often fragmented, which is precisely the scenario targeted by in-place completion.

4. Modalities and annotation structure

For each rendered view, ARSG-110K provides RGB images, camera intrinsics and extrinsics, per-pixel instance masks, and depth maps (Yin et al., 6 Apr 2026). For each scene and object instance, it provides complete 3D mesh models and 6DoF pose in scene coordinates, including translation and rotation matrices; scaling is also applied during generation and is implicitly part of the transformation (Yin et al., 6 Apr 2026).

The paper’s characterization of “high-fidelity 3D ground truth” is grounded in mesh-level supervision. The dataset is defined in terms of polygonal meshes rendered in Blender, with exact scene composition and object placement. During model training, these meshes can be transformed into occupancy grids at resolution $64^3$, SLAT sparse voxel features, coarse latent voxel grids at $16^3$ via a 3D VAE, and point-cloud or voxelized visible-region representations, but those latent forms are described as training-time internal representations rather than the primary dataset format (Yin et al., 6 Apr 2026).

At the object level, ARSG-110K supports supervision with complete mesh geometry, pose, and per-view instance masks. At the scene level, it supports supervision with the list of object instances, structural planes, HDR environment choice, and full camera parameters (Yin et al., 6 Apr 2026). The paper does not mention part-level or material-part labels, and explicitly centers asset-level supervision instead.

The exact released directory structure is not specified. The source description implies per-scene organization containing camera metadata, RGB images, depth maps, instance masks, object meshes, object transforms, and environment metadata (Yin et al., 6 Apr 2026). This suggests a scene-centric representation in which per-view and per-instance assets can be reconstructed without ambiguity.

5. Role in 3D-Fixer training and evaluation

ARSG-110K is the training backbone for 3D-Fixer (Yin et al., 6 Apr 2026). The method takes a single RGB scene image, an instance mask, fragmented visible point cloud, global 2D features, and camera parameters as input, and outputs a complete 3D asset aligned with its original pose in the scene. The dataset provides the ground-truth full mesh for each instance and the camera and depth information needed to construct the corresponding fragmented geometry used for conditioning.

A notable aspect of the training pipeline is depth mixing. The paper states that training does not use perfect depth alone. Instead, a depth estimator is sampled from {MoGe v2, VGGT, DepthAnything v2, Depth Pro}, producing $d_{\text{est}}$, which is then mixed with ground-truth depth $d_{\text{gt}}$ via

$$d = \alpha \cdot d_{\text{est}} + (1 - \alpha) \cdot d_{\text{gt}}, \quad \alpha \sim \mathcal{U}(0.0, 1.0).$$

During inference, $\alpha$ is set to $1.0$ (Yin et al., 6 Apr 2026). The mixed depth and camera parameters are used to reconstruct a point cloud, crop it with the instance mask, and obtain fragmented geometry $G_{\mathrm{frag}}$.

ARSG-110K supervises both stages of the coarse-to-fine generation process. In the Coarse Structure Completer, the visible point cloud is enclosed in an expanded box derived from its axis-aligned bounding box:

$$B_{\mathrm{vis}} \rightarrow C_{\mathrm{vis}}, \text{ side lengths } l_x, l_y, l_z,$$

with

$64^3$0

and the expanded box $64^3$1 centered at $64^3$2 with side length $64^3$3 (Yin et al., 6 Apr 2026). This is designed so that the true full object bounding box $64^3$4 fits inside $64^3$5 even under heavy occlusion. In the Fine Shape Refiner, supervision is derived from the SLAT representation computed from the ground-truth asset mesh (Yin et al., 6 Apr 2026).

The dataset is also central to the Occlusion-Robust Feature Alignment (ORFA) strategy. The feature alignment loss is defined as

$64^3$6

where $64^3$7 is a teacher latent feature from clean object images and $64^3$8 is the corresponding student feature from occluded scene-level inputs (Yin et al., 6 Apr 2026). ARSG-110K is necessary here because it provides both the occluded scene views and the clean per-asset views needed to train and align these representations.

The paper does not provide an explicit train/validation/test partition for ARSG-110K itself. It states that 3D-Fixer is trained on ARSG-110K and evaluated on the MIDI test set, the Gen3DSR test set, a new synthetic test set built from Toys4K assets using the same procedural composition pipeline, and real-world ScanNet scenes with MetaScenes ground truth (Yin et al., 6 Apr 2026). This suggests that the dataset’s primary role is pretraining and synthetic supervision rather than serving only as a closed benchmark.

6. Empirical significance and benchmark context

The 3D-Fixer paper attributes substantial generalization performance to training on ARSG-110K (Yin et al., 6 Apr 2026). On the MIDI test set, scene-level metrics improve from 0.080 / 50.19 to 0.069 / 78.67 for CD$64^3$9 / FS$16^3$0, and object-level metrics improve from 0.103 / 53.58 to 0.032 / 94.39 for CD$16^3$1 / FS$16^3$2 when comparing MIDI to 3D-Fixer (Yin et al., 6 Apr 2026). On the Gen3DSR test set, 3D-Fixer trained only on ARSG-110K reports 0.103 / 77.95 for CD$16^3$3 / FS$16^3$4, compared with 0.120 / 68.82 for Gen3DSR and 0.566 / 23.49 for MIDI (Yin et al., 6 Apr 2026).

On the Toys4K-based synthetic test set created with the same procedural pipeline but unseen assets, 3D-Fixer reports scene-level CD$16^3$5 / FS$16^3$6 of 0.159 / 68.82 and object-level CD$16^3$7 / FS$16^3$8 / IoU of 0.197 / 57.85 / 0.519 (Yin et al., 6 Apr 2026). On a ScanNet + MetaScenes subset, the method reports FS$16^3$9 of 61.58, compared with 50.19 for Gen3DSR and 37.10 for MIDI (Yin et al., 6 Apr 2026). These outcomes are presented as evidence that training solely on ARSG-110K yields strong transfer beyond the synthetic source domain.

The paper evaluates geometry with Chamfer Distance and F-Score and layout with 3D bounding-box IoU. It defines

$d_{\text{est}}$0

and F-Score at threshold $d_{\text{est}}$1 from precision and recall over point-set correspondence (Yin et al., 6 Apr 2026). For texture, the paper reports visible/unseen FID and CLIP metrics on the ARSG-style test set, with 3D-Fixer obtaining 43.52 / 46.25 for FID$d_{\text{est}}$2 / FID$d_{\text{est}}$3 and 89.82 / 88.51 for CLIP$d_{\text{est}}$4 / CLIP$d_{\text{est}}$5, compared with 102.72 / 119.22 and 80.70 / 77.71 for Gen3DSR (Yin et al., 6 Apr 2026).

These results do not by themselves define a benchmark protocol for ARSG-110K, but they indicate the kinds of tasks it is intended to support: single-image compositional 3D scene generation, in-place completion of partially visible assets, layout recovery, and photorealistic texture synthesis (Yin et al., 6 Apr 2026).

7. Limitations, interpretation, and future use

The paper does not frame ARSG-110K as a complete substitute for real-scene corpora. Several limitations are explicit or directly inferable from the dataset description (Yin et al., 6 Apr 2026). First, the dataset is fully synthetic, so a synthetic-to-real gap remains despite strong transfer to ScanNet-based evaluation. Second, scenes are built from a simple structural template—floor plus up to four walls—which does not reproduce the full architectural regularity of real buildings. Third, outdoor scenes are approximated by floor-only layouts rather than detailed natural or urban environments. Fourth, the dataset does not explicitly mention semantic category labels or part-level annotations, and its supervision is asset-centric rather than semantics-centric (Yin et al., 6 Apr 2026).

At the same time, the paper presents ARSG-110K as a foundation for future work in compositional 3D scene generation. Its combination of scale, exact asset geometry, precise transforms, and dense occlusion is explicitly intended to facilitate research on in-place completion, amodal reconstruction, layout-aware asset generation and retrieval, and occlusion handling in multi-instance scenes (Yin et al., 6 Apr 2026). A plausible implication is that ARSG-110K could serve as a pretraining substrate for scene-level 3D foundation models that require exact mesh supervision but must generalize to cluttered images.

The paper states that code and data will be publicly available at the 3D-Fixer project page and that the dataset and scene construction script will be made publicly available (Yin et al., 6 Apr 2026). No license is specified in the provided description. Because the asset pool incorporates multiple upstream repositories, downstream use will plausibly depend on those source licenses, although that point is not elaborated in the paper.

ARSG-110K is therefore best understood as a procedurally generated, asset-rich, scene-level dataset engineered for geometry-complete, pose-faithful, occlusion-heavy supervision. Its central distinguishing property is not only the reported scale of 110K scenes and 3M views, but the combination of that scale with exact per-object meshes, transformations, and render-derived multimodal annotations, all tailored to single-image 3D scene generation and in-place completion (Yin et al., 6 Apr 2026).