---
title: 'PartVerse: 3D Part-Annotated Dataset'
url: https://www.emergentmind.com/topics/partverse
type: topic
---

# PartVerse: 3D Part-Annotated Dataset

Searching arXiv for the cited papers and related work on PartVerse.
PartVerse is a large-scale, part-level 3D dataset introduced alongside CoPart, a part-aware diffusion framework for 3D generation. Constructed from Objaverse, it contains approximately 12,000 objects, approximately 91,000 textured parts, 175 object categories, and part-level textual descriptions. Its defining property is that objects are explicitly decomposed into semantically meaningful parts, each retaining geometry, texture, and natural-language description, so that generative models can learn over sets of contextual part latents rather than a single holistic object latent [2507.08772].

## 1. Origin, definition, and motivation

PartVerse was created to address a specific deficiency in the data regime for 3D-native generative modeling. Large corpora such as Objaverse are diverse and textured but are object-centric and do not provide clean, consistent part annotations. PartNet, by contrast, provides explicit part labels, but only for 24 categories, with poor or missing textures, and at a scale too limited for training high-capacity, generalizable diffusion models. PartVerse was designed as a 3D counterpart of Objaverse that preserves the scale and visual richness of web-scale 3D assets while introducing explicit, semantically meaningful part decompositions suitable for learning part-aware generative priors [2507.08772].

The dataset serves three purposes in the CoPart formulation. First, it enables part-level learning by providing thousands of textured parts across many categories. Second, because parts are derived from complete objects, it preserves part co-occurrence structure, allowing synchronized modeling of inter-part relationships such as chair seat, back, and legs. Third, it supports controllability through part-level text and geometry supervision, including local prompts and bounding-box conditions. In the CoPart paper, these requirements are not ancillary; they are the dataset-level preconditions for representing an object as a set of contextual part latents and learning a joint distribution over them [2507.08772].

A recurrent misconception is to treat PartVerse as a model or as a generic segmentation corpus. In the primary source, it is neither. It is the dataset backbone of CoPart: a large-scale, textured, part-annotated corpus specifically organized for diffusion-based 3D generation and editing [2507.08772].

## 2. Construction pipeline and annotation methodology

PartVerse is curated from Objaverse through a four-stage pipeline: raw data selection from Objaverse, automatic mesh segmentation to obtain initial part proposals, human post-annotation to correct and refine parts, and generation of part-level textual captions using a vision-language model. The stated objective is to restore the “part-by-part” design intent of human 3D artists in a consistent, machine-usable format [2507.08772].

The automated pre-labeling stage combines SAM-2 and Segment Any Mesh (Samesh). SAM-2 supplies semantic segmentation on rendered views, and Samesh lifts these 2D segmentations back to mesh faces. The pipeline explicitly favors over-segmentation rather than under-segmentation: fine fragments can be merged by annotators, whereas fused semantic parts are harder to correct. Two criteria are balanced during this stage: mesh connectivity, so that a part consists of connected mesh faces, and visual semantics, so that boundaries respect the object’s appearance in rendered RGB views. Parameter tuning is therefore biased toward finer segments before manual refinement [2507.08772].

Human post-annotation is performed in a Blender-based annotation platform. Annotators first filter out unsuitable assets, including objects that are too complex, noisy, or otherwise unfit for meaningful part splitting. They then merge over-segmented components, split under-segmented ones, preserve clear part semantics, and maximize symmetry in part distributions. After validation, each textured object is split into separate part meshes with textures preserved. The combination of automated over-segmentation and curated human correction is presented as the basis of PartVerse’s quality and consistency [2507.08772].

Part-level textual captions are generated from composite renders that show the whole object with the target part highlighted by a bounding box. For each part, the pipeline renders multi-view images of the full object and the isolated part, selects the view that maximizes visible overlap, and feeds the composite image to Mini-Gemini to describe local appearance, local shape, and the part’s relationship to the whole. These captions later function as part-level prompts for text-conditioned generation and editing [2507.08772].

## 3. Data representation and annotation schema

In PartVerse, each part $\mathcal{M}_p$ is a textured triangle mesh, but CoPart trains on encoded geometric and appearance representations rather than raw meshes. On the geometry side, the pipeline samples surface points and normals,
$$
\bm{P} \in \mathbb{R}^{S \times 3},\quad \bm{Q} \in \mathbb{R}^{S \times 3},\quad S = 4096,
$$
and encodes them with a 3D part VAE $\mathcal{E}_{3D}$ into latent tokens
$$
\bm{L}_{3D} = \mathcal{E}_{3D}(\bm{P}, \bm{Q}) \in \mathbb{R}^{T \times D}.
$$
The corresponding decoder $\mathcal{D}_{3D}$ predicts Flexicube parameters, enabling differentiable rendering with depth and normal supervision during VAE fine-tuning [2507.08772].

On the appearance side, each part is rendered into multi-view, part-centric images $\{\bm{O}_k\}_{k=1}^{v}$ and encoded with the PixArt-$\alpha$ image VAE:
$$
\bm{F}_k = \mathcal{E}_{2D}(\bm{O}_k) \in \mathbb{R}^{T \times D},\quad k = 1,\dots,v,
$$
with the appearance latent set collected as
$$
\bm{L}_{2D} = \{ \bm{F}_k \}_{k=1}^{v}.
$$
This yields a paired representation in which each part has geometric latent tokens $\bm{L}_{3D}$ and multi-view appearance latents $\bm{L}_{2D}$ [2507.08772].

The annotation schema is deliberately partial rather than exhaustive. PartVerse provides part–object grouping, preserved textures, and part-level text prompts. The CoPart paper states that 3D bounding boxes per part can be computed from the meshes and encoded as
$$
\bm{L}_{box}^{p} = \mathcal{E}_{3D}(\bm{P}_{box}^{p}, \bm{Q}_{box}^{p}),
$$
where $\bm{P}_{box}^{p}, \bm{Q}_{box}^{p}$ are sampled points and normals on the box surfaces. Explicit adjacency graphs, kinematic chains, and hierarchies are not annotated. However, the paper notes that articulation and layout information can be inferred from bounding boxes and the co-occurrence of parts inside the same original object [2507.08772].

This annotation profile places PartVerse between classical structured part datasets and purely object-level web corpora. It supplies enough semantics and geometry for joint part modeling, but it does not impose a fixed hierarchical ontology comparable to PartNet’s explicit trees.

## 4. Role in CoPart and part-aware 3D generation

CoPart represents an object as a set of part latents,
$$
\{ \bm{L}_{3D}^{p}, \bm{L}_{2D}^{p} \}_{p=1}^{N},
$$
and trains diffusion denoisers over these sets rather than over a single holistic latent. Each part is encoded, Gaussian noise is added at diffusion timestep $t$, and the 3D and 2D denoisers—initialized from pre-trained CraftsMan and PixArt-$\alpha$ respectively—predict the noise for the geometry and appearance branches. The training losses are averaged across the parts of each object, which means the model is trained directly on part sets sampled from PartVerse objects and therefore learns a joint distribution over part latents conditioned on text and box signals [2507.08772].

PartVerse is particularly important for CoPart’s mutual-guidance design. Inter-part synchronization contextualizes one part by the other parts of the same object through attention within each modality. Intra-part cross-modality guidance aligns geometry and appearance for the same part. A separate multi-view consistency mechanism regularizes appearance latents across views, and a global branch—effectively a shared “part 0”—interacts with all part latents to preserve object-level coherence. The paper’s argument is that such synchronized denoising is only learnable because PartVerse supplies consistently annotated multi-part objects with preserved textures and captions [2507.08772].

The dataset also underwrites CoPart’s conditioning interface. Part-level text prompts come from the VLM-generated captions, while part boxes can be encoded in 3D latent space and rendered as wireframes for the 2D branch via ControlNet. In the reported formulation, these conditions resolve part-order ambiguity and enable box-guided, text-guided part generation. The same learned structure supports part-level editing, articulated object generation through integration with CAGE, and mini-scene composition in which each scene object is treated as a “part” under a layout specified by bounding boxes [2507.08772].

Evaluation in the CoPart paper includes text-conditioned 3D generation measured by CLIP (N–T), CLIP (I–T), and ULIP-T, with part-aware variants using prompts that target sub-parts such as “a car wheel.” User studies comparing CoPart with Rodin and PartGen report preferences of 54.9% in whole-aware settings and 60.8% in part-aware settings. Within the paper’s framing, these results are evidence not only for the model architecture but for the efficacy of a part-centric training corpus [2507.08772].

## 5. Derived benchmarks, successors, and ecosystem extensions

PartVerse has subsequently been repurposed beyond CoPart. In 3D-PLOT-LLM, it is treated as a mesh-level part-annotated dataset on Objaverse, providing meshes with per-triangle part IDs and an unmodified semantic part vocabulary. That work derives PartVerse-QA, a bidirectional caption-to-slots and slot-to-caption benchmark aligned to a vocabulary of symbolic tokens $\langle\texttt{part\_k}\rangle$. The benchmark is built by aligning PartVerse meshes to 8192-point Objaverse point clouds through ICP plus Umeyama rigid alignment, partitioning each object into $K=16$ regions, mapping semantic parts to minimal slot sets with union IoU filtering at $\ge 0.5$, and instantiating question–answer pairs in both directions. The resulting benchmark contains 77,607 training pairs and 588 evaluation queries on object-disjoint held-out splits; on this benchmark, 3D-PLOT-LLM reports caption-to-slots Jaccard $0.459$, Exact-match $13.78\%$, and slot-to-caption GPT-4o judge $44.68$ [2606.19828].

A second extension is PartVerse-XL, introduced in FullPart as “an expanded and refined extension of PartVerse.” PartVerse-XL is described as the largest human-annotated 3D part dataset to date, with 40K objects, 320K manually refined parts, over 200 object categories, texture-preserving parts, and part-level textual captions. Its construction preserves the same broad logic as PartVerse—over-segment first, refine with human annotators, and generate part-aware language—but scales from Objaverse-XL and uses Qwen2.5-VL for caption generation. FullPart uses this successor dataset as its sole training resource for layout generation, per-part voxel generation, and textured mesh refinement [2510.26140].

CubePart extends the surrounding ecosystem still further by aggregating public sources including Objaverse, TexVerse, PartVerse, and PartVerse-XL with commercial libraries and internal game assets into a 462K-asset, 2.02M-part open-vocabulary part dataset. In that work, PartVerse and PartVerse-XL function as predecessors within a larger shift from fixed taxonomies to open-vocabulary, schema-aligned part naming and generation [2605.28763].

| Resource | Definition | Reported scale |
|---|---|---|
| PartVerse | Objaverse-based textured, part-annotated 3D dataset for CoPart | $\approx 12$K objects, $\approx 91$K parts, 175 categories |
| PartVerse-QA | Bidirectional part-QA benchmark adapted from PartVerse mesh annotations | 77,607 training pairs, 588 evaluation queries |
| PartVerse-XL | Expanded and refined extension of PartVerse for FullPart | 40K objects, 320K parts, over 200 categories |

These derivatives clarify the distinction between the original dataset and later artifacts. PartVerse is the foundational mesh-and-caption dataset; PartVerse-QA is a language benchmark derived from its annotations; PartVerse-XL is a larger successor dataset rather than a relabeling of the original resource.

## 6. Comparative position, limitations, and research significance

Within the landscape of 3D datasets, PartVerse sits between Objaverse and PartNet. From Objaverse it inherits large scale, textured and visually realistic meshes, and broad category coverage. From PartNet it inherits the idea of explicit per-part supervision and semantic guidance for part boundaries. The CoPart paper characterizes it as the first large-scale, textured, diverse, part-annotated dataset tailor-made for diffusion-based 3D generative modeling [2507.08772].

Its limitations are also explicit. Category imbalance and quality variance remain, particularly for objects with complex topology or ambiguous semantics. Some segmentation errors persist despite human refinement, and the cost of annotation prevents exhaustive manual correction for every asset. The dataset does not include explicit articulation annotations, joint parameters, or kinematic hierarchies; articulated generation in CoPart therefore relies on an external structure model, CAGE, rather than on PartVerse alone. The paper also notes downstream assembly issues such as mesh clipping when user-provided bounding boxes are poorly aligned, underscoring the importance of accurate box conditions in applications of PartVerse-trained models [2507.08772].

The broader research significance of PartVerse lies in the representational shift it supports. The dataset was built to enable movement from holistic object latents toward multi-part, context-aware representations in which representational capacity is distributed over semantically meaningful subcomponents. This shift makes part editing, swapping, recombination, and layout control natural operations rather than afterthoughts. A plausible implication is that PartVerse’s main legacy is not only a corpus of annotated parts, but a data model for 3D generation in which parts are treated as first-class statistical entities.

The future directions stated in the source are correspondingly structural: more categories and more objects per category, better and possibly hierarchical part taxonomies, explicit graph annotations such as parent–child links and articulation joints, temporal or dynamic part states, and integration with scene-level datasets. In that sense, PartVerse is best understood as both a dataset and an inflection point in part-aware 3D generative modeling: the initial resource that made contextual part latents operational at scale, and the precursor to later benchmarks and successors that extended part awareness into 3D multimodal language models and higher-resolution part generation [2507.08772].

Source: https://www.emergentmind.com/topics/partverse