3DEditVerse: Scalable 3D Editing Benchmark

• 3DEditVerse is a large-scale paired benchmark for localized 3D asset editing that provides precise before–after correspondences.
• It employs complementary pipelines—pose-driven geometric edits and foundation model-guided appearance edits—to ensure edit locality and multi-view consistency.
• The framework supports models like 3DEditFormer and Omni123, driving measurable improvements in 3D and 2D evaluation metrics.

Searching arXiv for the cited works to ground the article in papers. 3DEditVerse is a paired benchmark and data-generation framework for localized 3D asset editing, introduced to support scalable learning and evaluation of edits that modify geometry or appearance while preserving unedited structure and maintaining multi-view consistency. In the formulation associated with "Towards Scalable and Consistent 3D Editing" (Xia et al., 3 Oct 2025), 3DEditVerse provides before–after 3D correspondences rather than unrelated assets plus text, and is designed around three target properties: edit locality, multi-view consistency, and semantic alignment. In subsequent work, it also functions as supervised editing data for native 3D editors such as Omni123, which fine-tunes an autoregressive text-and-3D model on 3DEditVerse for instruction-based source-mesh-to-target-mesh editing (Ye et al., 2 Apr 2026).

1. Definition and scope

3DEditVerse is defined as a large-scale paired 3D editing benchmark in which each sample contains a source 3D asset and a target edited 3D asset related by a localized transformation rather than independent generation (Xia et al., 3 Oct 2025). The benchmark is explicitly motivated by the claim that 3D editing differs from 2D editing in requiring cross-view consistency, structural fidelity in preserved regions, fine-grained controllability, and practical scalability. Within that framing, paired supervision is treated as the critical missing ingredient: it specifies both what should change and what should remain fixed.

The benchmark is described as the largest paired 3D editing benchmark to date, with 116,309 training pairs and 1,500 curated test pairs (Xia et al., 3 Oct 2025). The same paper also reports an underlying corpus of about 118K paired 3D assets, decomposed into 54,000 structural or pose-driven assets and 64,123 appearance-oriented edits; the benchmark split numbers remain the operative training/test counts. The benchmark spans both geometric and appearance edits, and its intended use is twofold: supervised training of 3D editors and standardized quantitative evaluation of local 3D editing.

A central consequence of this design is that 3DEditVerse is not merely a prompt collection or an image-editing dataset lifted into 3D independently per sample. The source and edited targets are constructed to preserve correspondence, locality, and semantic harmony. This distinguishes it from datasets criticized in the same paper for lacking consistency or producing semantically awkward composites (Xia et al., 3 Oct 2025).

2. Construction pipelines

3DEditVerse is built through two complementary pipelines: a pose-driven geometric pipeline and a foundation model-guided appearance pipeline (Xia et al., 3 Oct 2025). Their complementarity is important because the first supplies naturally consistent structural variation at scale, while the second supplies semantically diverse instruction-conditioned edits.

2.1 Pose-driven geometric edits

The geometric branch is based on publicly available 3D characters and animation sequences from Mixamo (Xia et al., 3 Oct 2025). Candidate animation frames are sampled, rendered from a canonical view, embedded with DINOv2, and pruned by cosine similarity to remove near-duplicate poses. This yields 4,998 unique candidate poses. A set of 108 distinct characters is then paired with 500 poses randomly selected from that candidate pool, producing:

$108 \times 500 = 54{,}000$

unique 3D assets.

Before–after edit pairs are formed by associating different poses of the same character. The resulting edits are geometric in the sense that they alter pose, articulation, or part configuration while maintaining identity and full 3D consistency. The benchmark text does not formalize locality mathematically for this branch, but the intended locality follows from identity-preserving articulation rather than arbitrary object replacement (Xia et al., 3 Oct 2025).

2.2 Foundation model-guided appearance edits

The second branch is a text-to-image-to-3D lifting pipeline via a cascade of foundation models (Xia et al., 3 Oct 2025). It begins with the 4,585-word vocabulary from Tag2Text. For each vocabulary word, DeepSeek-R1 generates diverse descriptive prompts; the appendix example states that for “plane,” DeepSeek-R1 produces 20 descriptive prompts, each around 20 words. These prompts are passed to Flux.1-Dev to produce a centered source image on a pure white background.

Given the source image $I^\text{src}$, Qwen-VL generates an edit instruction $p^\text{edit}$, and Flux.1-Kontext executes the edit to obtain the target image $I^\text{tgt}$ (Xia et al., 3 Oct 2025). The appendix indicates add/remove/replace-part prompt templates with emphasis on substantial visible shape transformations, although the main text characterizes this branch overall as appearance-oriented. The pipeline is further augmented with samples from ShapeLLM-Omni rendered from Objaverse-XL, whose source image and prompt pairs are also edited with Flux.1-Kontext.

The resulting image pairs are then lifted into 3D with a consistency-preserving procedure rather than independently reconstructed source and target assets. This design reflects the paper’s claim that independent lifting causes geometric distortion and identity mismatch (Xia et al., 3 Oct 2025).

3. Locality and multi-view consistency mechanisms

The most technically distinctive part of 3DEditVerse is the appearance-edit lifting pipeline, which uses automatic region localization, multi-view mask propagation, 3D back-projection, and latent localized editing to ensure that source and target differ only where intended (Xia et al., 3 Oct 2025).

First, source and target images are separately lifted with Trellis to produce initial 3D assets $\hat{S}_{3D}^\text{src}$ and $\hat{S}_{3D}^\text{tgt}$. Then Qwen-VL is used as an open-set detector to output a 2D bounding box $B_{2D}$ for the edited region on a rendered source view. Across multiple rendered views, SAM2 segments and tracks the edited region, producing 2D masks $\{M_{2D}^i\}$.

These per-view masks are back-projected into 3D using the pinhole model. The paper gives the projection as:

$$\tilde{p}_i = K_i [R_i \mid t_i] v, \quad p_i = \left( \tfrac{\tilde{p}_{i,x}}{\tilde{p}_{i,z}}, \tfrac{\tilde{p}_{i,y}}{\tilde{p}_{i,z}} \right),$$

where $v=(x,y,z,1)^\top$ is a voxel, $I^\text{src}$0 is the intrinsic matrix, and $I^\text{src}$1 is the extrinsic matrix of view $I^\text{src}$2 (Xia et al., 3 Oct 2025). Multi-view support is accumulated over 70 views:

$I^\text{src}$3

and the final 3D mask is defined by thresholding support:

$I^\text{src}$4

This mask serves as the locality operator for latent editing. Using Repaint in Trellis, the paper injects noise into the source asset latent at timestep $I^\text{src}$5,

$I^\text{src}$6

and then fuses source and target latents via the 3D edit mask:

$I^\text{src}$7

Inside the mask, the target latent is followed; outside the mask, the source latent is preserved (Xia et al., 3 Oct 2025). This is the dataset’s main operational mechanism for encoding edit locality. Post-edit quality control then renders both the final edited asset and the initial target prediction from multiple views and filters samples with low mean DINOv2 cosine similarity, removing inconsistent or low-quality edits.

This pipeline suggests that 3DEditVerse is best understood not just as a static benchmark, but as a procedurally constructed paired-edit corpus whose supervision signal is already shaped by explicit geometric and consistency constraints.

4. Data content and benchmark structure

The released benchmark is described as containing paired original and edited 3D assets, and its construction additionally involves rendered images, edit prompts, 2D bounding boxes, tracked 2D masks, back-projected 3D masks, and camera parameters (Xia et al., 3 Oct 2025). The paper does not enumerate which intermediate fields are guaranteed to be part of the public release, but they are part of the generation and filtering process.

The benchmark is explicitly compared against three earlier resources (Xia et al., 3 Oct 2025):

Dataset	Train size	Test size
3D-Alpaca-Editing	52,532	—
CMD	40,000	50
Edit3D-Bench	—	300
3DEditVerse	116,309	1,500

The same comparison argues that 3DEditVerse uniquely combines train/test availability with good edit region quality, consistency, and harmony (Xia et al., 3 Oct 2025). Here, “consistency” refers to before–after correspondences derived from the same source identity rather than independent generation, while “harmony” refers to semantically plausible edited targets rather than concatenations of unrelated assets.

The test set is manually curated through human assessment, and the paper treats that curation as a benchmark-quality control stage (Xia et al., 3 Oct 2025). A plausible implication is that benchmark reliability depends not only on generation-scale automation but also on aggressive filtering and human validation.

5. Relationship to 3DEditFormer

3DEditVerse is introduced jointly with 3DEditFormer, a mask-free 3D editor built by extending Trellis (Xia et al., 3 Oct 2025). The benchmark’s role is not incidental: it supplies the paired supervision that allows the model to learn structural preservation and editable-region localization without requiring auxiliary 3D masks at inference.

3DEditFormer uses a Dual-Guidance Attention Block, Multi-Stage Feature Extraction, and Time-Adaptive Gating. In the $I^\text{src}$8-th transformer block, standard self-attention and two source-guided cross-attention branches are fused as

$I^\text{src}$9

$p^\text{edit}$0

$p^\text{edit}$1

The two feature families are extracted from the frozen Trellis model at different timesteps:

$p^\text{edit}$2

with $p^\text{edit}$3 and $p^\text{edit}$4 (Xia et al., 3 Oct 2025). The timestep-dependent gates are produced by

$p^\text{edit}$5

The generation objective follows Trellis’s Conditional Flow Matching formulation:

$p^\text{edit}$6

with

$p^\text{edit}$7

This architecture matters for understanding 3DEditVerse because the benchmark is the enabling supervision that teaches the model to separate preserved structure from edit-induced change. The claim is not merely that 3DEditVerse is large, but that its paired locality-preserving design makes mask-free structure-preserving editing learnable (Xia et al., 3 Oct 2025).

6. Evaluation, downstream use, and later relevance

On the 3DEditVerse test set, 3DEditFormer is evaluated using 3D metrics computed from 100,000 points sampled from predicted and ground-truth meshes—Chamfer Distance (CD), Normal Consistency (NC), and F1$p^\text{edit}$8—and 2D metrics computed from 10 fixed camera viewpoints—PSNR, SSIM, LPIPS, and DINO-I (Xia et al., 3 Oct 2025). Relative to EditP23, the paper reports large improvements on the full test set, including CD 46.19 → 13.84, NC 0.689 → 0.830, F1$p^\text{edit}$9 32.33 → 64.30, PSNR 18.32 → 24.40, SSIM 0.870 → 0.918, LPIPS 0.158 → 0.068, and DINO-I 0.785 → 0.963 (Xia et al., 3 Oct 2025). On the subset excluding character-animation samples, 3DEditFormer also surpasses VoxHammer on most reported metrics despite not using auxiliary 3D masks.

The benchmark further serves as editing supervision for Omni123, a native multimodal autoregressive model that unifies text, images, and 3D in a shared discrete-token framework (Ye et al., 2 Apr 2026). Omni123 includes a dedicated stage titled “Align to Instruction-Based 3D Editing” and uses 3DEditVerse with 116,309 training pairs and 1,500 curated test pairs for source-mesh-plus-instruction to target-mesh training (Ye et al., 2 Apr 2026). Its editing objective is

$I^\text{tgt}$0

where $I^\text{tgt}$1 is the source 3D token prefix and $I^\text{tgt}$2 is the instruction (Ye et al., 2 Apr 2026). This later use indicates that 3DEditVerse has already become more than a benchmark: it is an alignment dataset for native 3D editors.

A common misconception is that 3DEditVerse is a single editing model. In the literature represented here, it is primarily a dataset and benchmark (Xia et al., 3 Oct 2025), though it is closely associated with the paired model 3DEditFormer and later reused by Omni123 for instruction-based editing alignment (Ye et al., 2 Apr 2026). Another misconception is that it is solely an appearance-edit corpus; the benchmark explicitly includes both pose-driven geometric edits and foundation model-guided appearance edits (Xia et al., 3 Oct 2025).

A plausible broader implication is that 3DEditVerse marks a shift in 3D editing research from weakly supervised or mask-dependent pipelines toward paired, locality-aware supervised learning. The paper’s main claim is not simply scale, but that scale is coupled to correspondence fidelity, multi-view-consistent region localization, and edit harmony. In that sense, 3DEditVerse functions as both infrastructure and methodological statement: local 3D editing becomes learnable and benchmarkable when the data itself is constructed to encode what must be preserved, what may change, and how those changes remain coherent across views (Xia et al., 3 Oct 2025).