Nano3D-Edit-100k: 3D Editing Dataset

• Nano3D-Edit-100k is a large-scale dataset of 100k paired 3D editing examples generated via a training-free, mask-free pipeline using front-view renderings and automated text instructions.
• The dataset features balanced object categories and editing operations, leveraging novel region-aware merging techniques for high-fidelity, feed-forward 3D content generation.
• It provides rich metadata, multiple file formats, and standardized evaluation protocols, supporting reproducible research in interactive content creation, gaming, and robotics.

Nano3D-Edit-100k is a large-scale, fully automated, and mask-free dataset for paired 3D object editing, comprising over 100,000 high-quality source–target 3D editing examples. Designed to support feed-forward, high-fidelity, and controllable 3D editing research, it is constructed using the Nano3D framework, which operates without human-annotated masks or manual intervention. Each edit is generated through a training-free pipeline that leverages front-view renderings, automated text instructions, and novel merging techniques to ensure structural fidelity and category diversity across object types and editing operations (Ye et al., 16 Oct 2025).

1. Dataset Composition and Coverage

Nano3D-Edit-100k contains 100,000 editing examples. Each example consists of:

• A source 3D asset in the TRELLIS sparse voxel grid and structured latent (“SLAT”) form.
• A source front-view rendering at 256 × 256 px.
• A text editing instruction, generated using fixed templates (“add X to Y”, “remove X from Y”, “replace X with Z”).
• An edited front-view image, produced by Nano-Banana or Flux-Kontext editors.
• A target (edited) 3D asset, stored as a SLAT file and optionally as a GLB/OBJ mesh (via FlexiCube).
• A JSON metadata file providing category, camera intrinsics and extrinsics, instruction, runtime-computed merging mask, and quality flags.

The dataset balances object categories across ten high-level classes, with the assignment automated by Qwen2.5-VL. Editing operations (addition, removal, replacement) are equally represented. The following table summarizes category and operation distributions:

Category	Example Count	Percentage
Human	20,755	20.75%
Weapon	11,021	11.03%
Furniture	10,442	10.45%
Personal Item	10,277	10.28%
Animal	10,186	10.19%
Vehicle	9,376	9.38%
Building	9,005	8.97%
Plant	4,441	4.45%
Electronic Device	5,283	5.29%
Else	6,593	6.60%

Each operation—addition, removal, replacement—has approximately 33,000 examples.

2. Automated Data Generation Pipeline

The construction of Nano3D-Edit-100k proceeds via a five-stage, training-free automated pipeline:

1. Sample a front-view image $I_{src}$ from a large image pool.
2. Generate a textual edit instruction from $I_{src}$ using a vision-LLM (VLM) like Qwen2.5-VL, choosing randomly among three fixed prompt templates.
3. Reconstruct the source 3D asset via TRELLIS, voxelizing at a $64^3$ grid and inferring the SLAT representation.
4. Produce an edited front-view image using the pretrained Nano-Banana or Flux-Kontext image editor, conditioned on the instruction.
5. Generate the edited 3D asset using Nano3D.edit(), which applies FlowEdit and region-aware Voxel/Slat-Merge strategies.
6. Apply quality filtering to images and meshes; only passing examples are retained in the dataset.

Editing is performed without masks; all region selections are determined algorithmically based on object structure changes.

3. Region-Aware Merging and Editing Methodology

Editing in Nano3D follows these principal stages:

• FlowEdit: Source ($I_{src}$) and target ($I_{tgt}$) renderings condition the TRELLIS model. Rather than standard flow-based sampling, Nano3D integrates source and target velocity fields along ODE trajectories, interpolating with a mixing schedule $\alpha_t$, yielding an edited voxel grid $V_{edited}$.
• Voxel-Merge: Compute a voxel-wise XOR ($g(i) = s_{src}(i) \oplus s_{tgt}(i)$) on sparse occupancies $s_{src}, s_{tgt} \in \{0,1\}^{64^3}$. Connected components $g_j$ are identified in 6/18/26-neighborhoods and filtered by a threshold $I_{src}$0 voxels. The merging mask $I_{src}$1 selects components above this threshold:

$I_{src}$2

Final voxel merging is performed as:

$I_{src}$3

with $I_{src}$4 if $I_{src}$5.

• Latent-Merge (SLAT): The merged voxel grid and edited image are used in Stage 2 TRELLIS to generate new structured latents $I_{src}$6. Merging in latent space yields:

$I_{src}$7

Decoding the merged SLAT with the TRELLIS VAE ensures globally consistent 3D asset generation.

No human-drawn masks are included; mask computation is internal and not required at test time.

4. Dataset Statistics, Formatting, and Splits

Nano3D-Edit-100k provides the following file types per example:

• SLAT file in PyTorch (.pt) format: contains the sparse voxel grid and array of local latent vectors $I_{src}$8.
• GLB/OBJ meshes (optional): produced via the FlexiCube converter for use in downstream applications.
• Front-view images: both source and edited renderings, 256 × 256 px in .png format.
• Metadata JSON: includes object category, operation type, edit instruction, camera intrinsics/extrinsics, merging mask, voxel occupation count, and quality flag.

Dataset characteristics at release:

• Polygon count (GLB): median ≈ 30,000 faces, range 5,000–200,000.
• Median voxel occupancy: 1.2 k active voxels per 3D object (on $I_{src}$9 grid).
• Front-view resolution: 256² px; additional multi-view renderings at 128² px.
• Stratified split recommended: 80% train (≥ 80k), 10% validation (10k), 10% test (10k), with preservation of category and operation balance.

5. Evaluation Protocols and Baseline Metrics

Nano3D-Edit-100k includes standardized evaluation protocols and public baseline results. Each 3D asset is rendered into $64^3$0 multi-view images (e.g., 8 azimuthal views). Recommended metrics:

• Chamfer Distance (CD): Quantifies structure preservation in non-edited regions; calculated as

$64^3$1

where $64^3$2 are point samples from unedited voxels.

• Target Semantic Alignment (DINO-I score): Compares edited image semantics using DINO-I between rendered edited views and the edited image.
• Generation Fidelity & Diversity (FID score): Computes Fréchet Inception Distance between rendered edited multiview images and ground-truth.

Reported baseline metrics (100 randomly chosen examples):

Method	CD	DINO-I	FID
Tailor3D	0.037	0.759	140.93
Vox-E	—	0.782	117.12
TRELLIS	0.019	0.901	49.57
Nano3D	0.013	0.950	27.85

Semantic alignment at the dataset level uses CLIPScore and ViLT R-Precision. Nano3D-Edit-100k achieves CLIPScore 39.71 and ViLT R-Precision (R@5: 45.3%, R@10: 52.4%), outperforming the prior 3D-Alpaca dataset (28.42; 33.6%, 40.2%).

6. Research Implications and Applications

Nano3D-Edit-100k addresses challenges in paired 3D editing supervision, absence of large-scale high-fidelity 3D edit pairs, and the need for mask-free pipelines. By providing diverse, balanced, and high-throughput data, it enables the systematic evaluation and development of fully feed-forward 3D editing models. Its mask-free, automated approach makes it suitable for advancing real-time and controllable 3D content generation in domains such as interactive content creation, animation, gaming, and robotics. A plausible implication is that the dataset will facilitate the transition to third-stage 3D editing: learning feed-forward, high-recall, real-time systems trained on large-scale, richly paired 3D supervision (Ye et al., 16 Oct 2025). All code for data loading, rendering, and metric computation is announced for release alongside the data, fostering reproducibility and benchmarking in the 3D editing community.