← 返回 2026-03-20

DreamPartGen:基于协作潜在去噪的语义部分级3D生成 DreamPartGen: Semantically Grounded Part-Level 3D Generation via Collaborative Latent Denoising

Tianjiao Yu, Xinzhuo Li, Muntasir Wahed, Jerry Xiong, Yifan Shen, Ying Shen, Ismini Lourentzou 📅 2026-03-19 👍 1 2026-07-13 08:36
扩散模型 文本到3D生成 结构化表示 语义对齐 部分感知生成

提出DPL和RSL协同去噪框架,实现语义驱动的部分级文本到3D生成

前置知识

扩散模型

扩散模型通过正向加噪和逆向去噪过程生成数据。在训练中,模型学习从纯噪声逐步恢复出真实数据;推理时,从随机噪声开始迭代去噪得到生成结果。Diffusion Transformer (DiT) 是一种将Transformer与扩散过程结合的架构,通过交叉注意力实现条件控制。本文采用两阶段训练:第一阶段优化部分潜在表示的扩散目标,第二阶段联合微调部分级和对象级同步模块。

本文核心创新是在扩散过程中引入语义协作,需要理解扩散的噪声调度、SNR加权课程等机制才能把握DPL-RSL如何协同去噪

文本到3D生成

文本到3D生成旨在根据自然语言描述创建3D模型。早期方法如DreamFusion使用分数蒸馏采样(SDS),将预训练的2D扩散模型作为3D优化的先验。近期改进包括可微渲染与显式3D表示结合(如Gaussian splatting在DreamGaussian中)、体素或网格参数化(如CLAY),以及混合自回归架构(如Trellis)。这些方法虽能生成单对象,但往往在保真度和多视角一致性上存在问题,且缺乏部分级建模。

本文要在文本到3D生成中加入部分级语义对齐,需要理解SDS等现有方法为何无法处理部分关系,以及DPL如何超越单对象整体生成范式

部分感知3D表示

部分感知方法将3D对象分解为语义上有意义的子部分(如椅子有座椅、靠背、扶手、腿)。PartNet提供分层部分标注,PartVerse扩展了更多类别。现有方法如Part123和Salad关注部分分割和组装,PartGen利用部分分解进行生成建模,CoPart引入部分级2D和3D潜在的双重先验,PartGS和Part2GS采用Gaussian splatting进行关节部分感知生成。但这些方法主要依赖几何信号(如边界框),语言指导未被充分探索。

本文的DPL设计建立在这些结构化潜在表示基础上,但关键区别在于引入持久化的语义关系,需要理解现有几何导向方法的局限性

潜在空间表示

潜在表示将高维数据(如图像、3D形状)压缩到低维空间以便神经网络处理。3D VAE将点云和法线编码为潜在序列 $L^{3D} \in \mathbb{R}^{T_{3D} \times d}$,图像VAE将渲染视图编码为 $L^{2D} \in \mathbb{R}^{T_{2D} \times d}$。这些潜在序列通过注意力机制交互,捕捉跨模态关系。本文的DPL创新在于为每个部分分配可学习的标识嵌入 $e_i \in \mathbb{R}^d$,确保部分在去噪时步间保持可追踪性。

本文核心是DPL的联合3D/2D潜在表示设计,需要理解潜在空间的编码、解码和注意力交互才能把握几何-外观协同的原理

研究动机

现有文本到3D方法主要操作整体潜在表示,将几何、外观和语义纠缠在一起,没有显式表示部分身份或部分间关系。例如,描述战斗机时用户可能指定"机翼下挂载多枚导弹",现有方法无法理解机翼和导弹之间的支撑或附着关系。近期部分感知方法虽然通过部分分割或边界框引导引入了分解,但它们对分割噪声脆弱,难以跨多样类别和提示扩展。更重要的是,许多基于部分的框架仍将部分视为几何上孤立的单元,不将部分间关系建模为显式变量,语言基本不可操作。这导致生成的3D对象经常出现部件分离、错位、功能失效等问题,如车轮悬浮在底盘外、机翼与机身断开连接、机械部件空间漂移等。

本文的目标是本文提出DreamPartGen框架,目标是实现语义驱动的、部分感知的文本到3D生成。具体而言,模型应该能够:(1)根据文本描述生成结构一致、功能合理的3D对象;(2)显式建模部分间的关系(如支撑、附着、对称、铰接等),这些关系源于自然语言;(3)保持部分身份在去噪过程中的一致性,避免槽位交换;(4)支持下游应用如部分编辑、关节对象生成和小场景合成。最终目标是让3D生成像人类理解对象一样——将对象视为有意义部分的组合,而不仅仅是一团纠缠的几何体。

与已有工作不同的是,本文的独特切入角度是引入持久化的、语言衍生的关系语义潜在表示(RSLs),这些潜在表示在整个去噪过程中保持活跃,而不仅仅作为一次性条件。与现有方法不同,DreamPartGen将部分级几何/外观(DPLs)与关系语义(RSLs)通过同步去噪进行联合建模,强制执行相互的几何-语义一致性。RSLs分为全局关系令牌 $S^{glb}$(作为固定结构条件)和局部语义令牌 $S^{loc}$(被噪声化并与部分潜在表示一同去噪),这样在每一去噪步都能提供语义细化和全局规划信号。此外,本文构建了PartRel3D数据集,通过30万个规范化功能与空间三元组,为语言驱动的关系监督提供大规模显式信号。

核心方法

DreamPartGen的整体思路是将基于部分的3D生成建模为两种互补潜在表示之间的语义驱动协作扩散过程:Duplex Part Latents (DPLs)编码单个部分的几何和外观,Relational Semantic Latents (RSLs)提供语言衍生的细节和全局结构。在训练中,模型首先通过部分级扩散目标优化DPL,然后在SNR加权课程下联合微调完整模型,包括部分级和对象级同步。推理时,输入提示被编码为局部语义令牌 $S^{loc}$ 和全局规划器令牌 $S^{glb}$,通过部分级和对象级同步联合去噪DPLs和RSLs,最终解码为部分网格并组装成对象。整个框架的关键是在去噪过程中保持关系语义的持久性和一致性。

核心创新点是引入持久化的、语言衍生的关系语义潜在表示,这些表示在去噪过程中持续存在,而不仅仅作为一次性条件。具体而言,DPLs通过两个互补元素建模每个部分:3D令牌序列 $L^{3D}_i$(通过3D VAE编码表面点和法线得到)和2D令牌序列 $L^{2D}_i$(通过预训练图像VAE编码多视图渲染得到),加上可学习的部分标识嵌入 $e_i$。RSLs包括全局关系令牌 $S^{glb} = \{s^{glb}_{ij,\rho}\}_{(i,j,\rho) \in \mathcal{R}}$(从规范化三元组$(i,j,\rho)$投影得到)和局部语义令牌 $S^{loc} = \{s^{loc}_m\}_{m=1}^K$(从部分级属性短语编码得到)。在去噪过程中,通过部分内同步(几何-外观对齐)和部分间同步(跨部分上下文传播+关系指导)实现DPLs和RSLs的协同演化。

方法步骤详情

方法步骤完整描述如下:\n\n步骤1:DPL实例化。对于每个部分网格 $p_i$,采样表面点及其法线,通过3D VAE编码器得到几何潜在序列 $L^{3D}_i \in \mathbb{R}^{T_{3D} \times d}$。同时,从多个视角渲染该部分,通过预训练图像VAE得到外观潜在序列 $L^{2D}_i \in \mathbb{R}^{T_{2D} \times d}$。为每个部分分配可学习的标识嵌入 $e_i \in \mathbb{R}^d$,确保部分在去噪时步间的可追踪性。\n\n步骤2:RSL实例化。使用冻结的文本编码器 $\mathcal{E}_{text}$ 将提取的关系/属性短语编码,然后通过学习的投影 $\phi_{text}$ 得到全局关系令牌 $S^{glb}$ 和局部语义令牌 $S^{loc}$。全局关系令牌从规范化的功能三元组(如$\texttt{support, attach, hinge, symmetry}$)和空间三元组(如$\texttt{above, touching, symmetric-with}$)得到,每个令牌编码两个部分如何语义相关。局部语义令牌从细粒度属性短语(如"metallic blade", "wooden handle")编码,用于细化和外观。\n\n步骤3:训练阶段一。优化3D和2D DPLs的扩散目标,在RSLs的语义条件下。对于时间步 $t \sim \mathcal{U}\{1, \ldots, T\}$ 和噪声 $\epsilon \sim \mathcal{N}(0, I)$,每部分扩散损失为 $\mathcal{L}^{3D}_{diff} = \mathbb{E}_{t,\epsilon}[\|\epsilon - \hat{\epsilon}_\theta(L^{3D}_{i,t}, L^{2D}_{i,t}, S^{glb}, S^{loc,t}, t)\|^2]$,$\mathcal{L}^{2D}_{diff}$ 类似。\n\n步骤4:训练阶段二。联合微调3D和2D部分去噪器及同步模块,使用SNR加权课程 $w_{syn}(t) = \frac{\text{SNR}(t)}{1+\text{SNR}(t)}$,其中 $\text{SNR}(t) = \alpha_t^2/\sigma_t^2$,逐步将焦点从保真去噪转向关系对齐。\n\n步骤5:推理。编码输入提示得到 $S^{loc}$ 和 $S^{glb}$。当显式三元组可用时,将其编码为 $S^{glb}$;否则默认为仅提示条件。采样高斯噪声初始化几何和外观流 $\{L^{3D}_{T,i}, L^{2D}_{T,i}\}_{i=1}^N$,应用相同的前向噪声过程得到 $S^{loc,T}$。从时间步 $T$ 到 $1$,使用部分级和对象级同步模块联合去噪 $\{L^{3D}_{i,t}, L^{2D}_{i,t}\}_{i=1}^N$ 和 $S^{loc,t}$,同时在持久 $S^{glb}$ 条件下。去噪后,将最终几何潜在 $\{L^{3D}_{0,i}\}_{i=1}^N$ 通过3D VAE解码器解码得到部分网格,组装成对象。

技术新颖性

技术新颖性体现在三个方面:(1)持久化语义关系建模。与一次性文本条件不同,RSLs在每一步去噪时注入,实现迭代语义细化。全局关系令牌 $S^{glb}$ 作为固定结构条件持续存在,局部语义令牌 $S^{loc}$ 与部分潜在表示一同去噪,细化和外观在语义约束下。(2)协作扩散架构。DPLs和RSLs通过部分内同步(公式3)和部分间同步(公式4)协同演化,强制几何-外观对齐和全局关系一致性。部分内同步对齐几何和外观,部分间同步通过直接消息传递和关系指导传播上下文。(3)大规模关系监督。PartRel3D数据集提供300K规范化的功能和空间三元组,跨越175个对象类别,平均每个对象有8.2个部分和27个部分间关系,提供密集的结构监督。这是首个显式语言驱动的关系监督数据集。

DreamPartGen connects part-level geometry and appearance with language-driven relational semantics, providing precise control over how parts are modified, arranged, and contextualized. This unified representation enables a wide range of downstream applications, including fine-grained part editing, articulated object generation, and mini-scene synthesis.
Figure 1: DreamPartGen connects part-level geometry and appearance with language-driven relational semantics, providing precise control over how parts are modified, arranged, and contextualized. This unified representation enables a wide range of downstream applications, including fine-grained part editing, articulated object generation, and mini-scene synthesis.
DreamPartGen Overview. DreamPartGen performs text-guided 3D generation by jointly denoising geometry, appearance, and relational semantics. Each object is decomposed into parts represented as Duplex Part Latents (DPLs) from 3D and 2D encoders, while Relational Semantic Latents (RSLs) encode text-derived details and global structure. Through intra-part (geometry–appearance alignment) and inter-part (relational planning via language) synchronization, DreamPartGen co-denoises DPLs and RSLs, enabling semantically grounded reconstruction of coherent part-aware 3D objects.
Figure 2: DreamPartGen Overview. DreamPartGen performs text-guided 3D generation by jointly denoising geometry, appearance, and relational semantics. Each object is decomposed into parts represented as Duplex Part Latents (DPLs) from 3D and 2D encoders, while Relational Semantic Latents (RSLs) encode text-derived details and global structure. Through intra-part (geometry–appearance alignment) and inter-part (relational planning via language) synchronization, DreamPartGen co-denoises DPLs and RSLs, enabling semantically grounded reconstruction of coherent part-aware 3D objects.
PartRel3D dataset overview of structured functional and spatial triplets for fine-grained inter-part semantic supervision.
Figure 3: PartRel3D dataset overview of structured functional and spatial triplets for fine-grained inter-part semantic supervision.

实验结果

核心发现包括:在几何保真度方面,DreamPartGen在所有基准上实现最低的CD和EMD,平均比之前方法降低53%的CD和33%的EMD。例如在PartRel3D上,CD从0.371降至0.081(降低78.2%),EMD从1.474降至0.412(降低72.0%)。在文本-形状对齐方面,在对象级别提升≥20%(CLIP(I-T)从0.207增至0.264,提升27.5%;ULIP-T从0.162增至0.197,提升21.6%),在部分级别提升≥35%(CLIP(I-T)从0.145增至0.200,提升37.9%;ULIP-T从0.109增至0.153,提升40.4%)。在部分独立性方面,平均IoU降低27.2%(从0.518降至0.377),表明更强的非交叉部分生成能力。在感知质量方面,render-FID/KID和P-FID/P-KID在所有数据集上均最佳,例如在PartRel3D上,P-FID从0.7517降至0.6921。在消融实验中,移除全局关系令牌(✗Sglb)导致CD增加275.1%,EMD增加101.4%,部分重叠(IoU)增加176.9%,ULIP-T降低46.8%;移除局部语义令牌(✗Sloc)导致CD增加647.6%,EMD增加438.6%,IoU增加207.5%,ULIP-T降低43.7%;移除部分标识符导致IoU增加106.6%,CD/EMD增加121.7%/91.0%,ULIP-T降低42.4%。在少部分和未见关系泛化方面,OOD-parts设置下Render-FID增加∆0.629(相比PartCrafter的∆1.072),OOD-rel设置下增加∆0.848(相比PartCrafter的∆1.449),且ULIP-T仅下降∆0.012-0.014。在推理效率方面,对象级生成耗时109秒,部分级生成45秒,3D场景生成52秒,在保持竞争性能的同时保持效率。

Quantitative evaluation on 3D object generation. Quantitative comparison with state-of-the-art methods on Objaverse, ShapeNet, ABO, and Partverse. Highlighted best and second-best results.
Table 1: Quantitative evaluation on 3D object generation. Quantitative comparison with state-of-the-art methods on Objaverse, ShapeNet, ABO, and Partverse. Highlighted best and second-best results.
Perceptual evaluation on part-level 3D object generation. r-FID/r-KID refers to render-FID/render-KID computed from multi-view renderings. P-FID/P-KID computed in PointNet++ feature space. Highlighted best and second best.
Table 4: Perceptual evaluation on part-level 3D object generation. r-FID/r-KID refers to render-FID/render-KID computed from multi-view renderings. P-FID/P-KID computed in PointNet++ feature space. Highlighted best and second best.
Quantitative evaluation with different input settings. We evaluate DreamPartGen with different combinations of input information. Among single-source inputs, spatial triplets (ST) offer the greatest improvements, while functional triplets (FT) alone contribute weaker geometric constraints. Combining Text+Image+FT+ST yields the best overall performance across all metrics, highlighting the complementary benefits of language-grounded functional and spatial relations.
Table 5: Quantitative evaluation with different input settings. We evaluate DreamPartGen with different combinations of input information. Among single-source inputs, spatial triplets (ST) offer the greatest improvements, while functional triplets (FT) alone contribute weaker geometric constraints. Combining Text+Image+FT+ST yields the best overall performance across all metrics, highlighting the complementary benefits of language-grounded functional and spatial relations.
Relation parsing robustness at inference. We compare three inference conditions for constructing relational triplets during test time: (1) with additional VLM (Qwen2.5-VL & ChatGPT-5), (2) oracle relations with ground-truth triplets, and (3) prompt-only conditioning. For the prompt-only row, we report the gap to the best-performing variant in each metric in parentheses. Highlighted best and second-best.
Table 6: Relation parsing robustness at inference. We compare three inference conditions for constructing relational triplets during test time: (1) with additional VLM (Qwen2.5-VL & ChatGPT-5), (2) oracle relations with ground-truth triplets, and (3) prompt-only conditioning. For the prompt-only row, we report the gap to the best-performing variant in each metric in parentheses. Highlighted best and second-best.
Quantitative evaluation on part-level 3D generation. We compare methods that explicitly generate part meshes on part-annotated datasets. Best and second-best results are highlighted.
Table 7: Quantitative evaluation on part-level 3D generation. We compare methods that explicitly generate part meshes on part-annotated datasets. Best and second-best results are highlighted.
Ablation on the number of local RSL tokens Km. We vary the number of semantic controller tokens used for local relational-semantic guidance. Highlighted best and second-best results.
Table 8: Ablation on the number of local RSL tokens Km. We vary the number of semantic controller tokens used for local relational-semantic guidance. Highlighted best and second-best results.
Robustness to rare parts and held-out relations. We evaluate in-distribution (ID) test data and two out-of-distribution (OOD) splits: OOD-parts (tail part labels) and OOD-rel (held-out relation predicates). We report absolute scores and the change relative to ID in parentheses (∆).
Table 9: Robustness to rare parts and held-out relations. We evaluate in-distribution (ID) test data and two out-of-distribution (OOD) splits: OOD-parts (tail part labels) and OOD-rel (held-out relation predicates). We report absolute scores and the change relative to ID in parentheses (∆).
Inference-time Comparison. We report per-sample inference latency. Methods are grouped by task type (object-level, part-level, or scene-level generation); timings should be interpreted within each row. Best time is highlighted.
Table 10: Inference-time Comparison. We report per-sample inference latency. Methods are grouped by task type (object-level, part-level, or scene-level generation); timings should be interpreted within each row. Best time is highlighted.
Evaluation on 3D Object-Composed Scene Generation. We report results on 3D-Front and an Occluded subset. We report CD↓, F-score↑, and IoU↓, together with inference runtime↓ per scene. Highlighted best and second-best.
Table 11: Evaluation on 3D Object-Composed Scene Generation. We report results on 3D-Front and an Occluded subset. We report CD↓, F-score↑, and IoU↓, together with inference runtime↓ per scene. Highlighted best and second-best.
Qualitative comparison on part-level 3D generation. Across diverse object categories, DreamPartGen yields the most faithful decompositions, preserving clear part boundaries, correct topology, and consistent spatial alignment. Baselines frequently exhibit assembly failures such as missing or detached parts (e.g., wing/head), spatial drift of small components (e.g., wheels/mechanical parts floating off the chassis), and unstable attachments that create surface tearing or holes around high-contact regions (neck, torso, shoulders, limb joints).
Figure 4: Qualitative comparison on part-level 3D generation. Across diverse object categories, DreamPartGen yields the most faithful decompositions, preserving clear part boundaries, correct topology, and consistent spatial alignment. Baselines frequently exhibit assembly failures such as missing or detached parts (e.g., wing/head), spatial drift of small components (e.g., wheels/mechanical parts floating off the chassis), and unstable attachments that create surface tearing or holes around high-contact regions (neck, torso, shoulders, limb joints).
Ablation on the co-denoising process with local semantic tokens Sloc. The condition-only baseline (−Sloc) yields coarse geometry and weak semantic coherence between parts.
Figure 5: Ablation on the co-denoising process with local semantic tokens Sloc. The condition-only baseline (−Sloc) yields coarse geometry and weak semantic coherence between parts.
Ablation on the global relational tokens Sglb. The model exhibits part-level misalignment and spatial drift without Sglb.
Figure 6: Ablation on the global relational tokens Sglb. The model exhibits part-level misalignment and spatial drift without Sglb.
Mini-scene generation. Given a scene-level description, DreamPartGen can generate a complete, coherent 3D scene with physically plausible spatial layouts, capturing object geometry and fine-grained object- and part-level relations.
Figure 7: Mini-scene generation. Given a scene-level description, DreamPartGen can generate a complete, coherent 3D scene with physically plausible spatial layouts, capturing object geometry and fine-grained object- and part-level relations.
Part editing. From a source object (left), DreamPartGen can execute relational edit prompts to place accessories on top of or around the head, while preserving geometry and spatial consistency.
Figure 8: Part editing. From a source object (left), DreamPartGen can execute relational edit prompts to place accessories on top of or around the head, while preserving geometry and spatial consistency.
Mini-scene Generation Results.
Figure 10: Mini-scene Generation Results.
查看结构化数据
任务指标本文基线提升
3D对象生成(几何保真度) Chamfer Distance (CD) ↓ 0.081 (PartRel3D) PartCrafter: 0.371 降低78.2%
3D对象生成(几何保真度) Earth Mover's Distance (EMD) ↓ 0.412 (PartRel3D) PartCrafter: 1.474 降低72.0%
3D对象生成(部分独立性) 平均配对IoU ↓ 0.212 (PartRel3D) HoloPart: 0.716 降低70.4%
3D对象生成(文本-形状对齐) CLIP(I-T) ↑ 0.264 (Partverse, 对象级) PartCrafter: 0.207 提升27.5%
3D对象生成(文本-形状对齐) ULIP-T ↑ 0.197 (Partverse, 对象级) PartCrafter: 0.162 提升21.6%
3D场景生成 CD ↓ 0.1495 (3D-Front) MIDI: 0.1602 降低6.7%
3D场景生成 F-Score ↑ 0.8146 (3D-Front) MIDI: 0.7931 提升2.7%

局限与改进

作者承认的局限性包括:(1)对关系解析器的依赖。虽然prompt-only推理保持竞争性,但使用VLM解析关系可以获得额外性能提升,这引入了对外部模型的依赖。(2)复杂场景的可扩展性。当前方法适用于小场景(几个对象),但扩展到大型复杂场景可能需要更高效的关系建模和注意力机制。(3)部分分解的一致性。虽然方法在OOD部分上表现良好,但仍然依赖于训练时的部分分解一致性,对于极不规范的分解可能仍然脆弱。\n\n自己观察到的局限性包括:(1)三元组规范的刚性。将自由形式的关系短语规范化为有限词汇表可能丢失细微差别,例如"slightly above"和"well above"都被映射为"above",这可能影响细粒度空间控制。(2)动态关系的处理缺失。当前RSLs建模静态关系,但某些应用可能需要动态关系(如"打开"门、"折叠"椅子),这需要更复杂的时序建模。(3)部分层次的缺失。PartNet提供分层部分标注,但DPLs是扁平的,无法显式建模部分-子部分关系,这可能限制对复杂层次结构的生成能力。(4)计算复杂度。虽然推理时间合理,但RSLs的双级同步增加了计算开销,对于实时应用可能需要优化。

独立分析的弱点

独立分析的弱点及改进方向:\n\n(1)三元组粒度粗的问题。当前关系谓词(如support、attach)是高层次的抽象,不直接约束几何。改进方向:引入细粒度几何约束,如接触点法向量对齐、关节轴参数等,将功能关系与具体几何参数关联。\n\n(2)VLM解析噪声的问题。使用VLM从文本或图像提取三元组可能引入错误或歧义。改进方向:训练端到端的关系解析模块,直接从原始文本或图像学习关系表示,避免显式三元组提取的误差传播。\n\n(3)缺乏物理仿真验证的问题。虽然生成的关系语义合理,但没有物理引擎验证功能正确性(如可移动关节)。改进方向:集成物理仿真器(如PyBullet、MuJoCo),在生成后验证和修正物理合理性,或训练时引入物理一致性损失。\n\n(4)部分层次建模缺失的问题。复杂对象(如人体、车辆)有层次结构(身体-手臂-手-手指),扁平的DPLs无法显式建模。改进方向:扩展DPLs为树状结构,引入部分-子部分注意力,分层同步几何-外观-语义。\n\n(5)时序关系和动态生成缺失的问题。当前只处理静态3D对象,无法生成可动画化的内容。改进方向:引入时序RSLs,建模动作和状态转移(如"打开"、"旋转"),支持生成可动画的3D资产。

未来方向

未来研究方向包括:\n\n作者提出的方向:(1)扩展到更复杂的具身或交互式设置,如机器人抓取场景中需要精确部分-部分交互推理;(2)探索结构化部分表示在更复杂语义任务中的作用,如问答、推理、规划。\n\n基于成果可延伸的方向:(1)多模态输入增强。当前主要使用文本,可以扩展到图像、草图、语音等多模态输入,实现更灵活的3D创作。(2)跨域迁移。将RSLs框架迁移到其他生成任务,如文本到4D视频、文本到场景图、文本到机器人技能等。(3)自监督学习。利用大规模无标注3D数据,通过自监督学习预训练部分级表示,减少对标注数据的依赖。(4)实时优化。开发更高效的注意力机制(如线性注意力、稀疏注意力),实现实时或近实时的3D生成。(5)用户交互设计。构建交互式编辑界面,用户可以实时修改部分关系、添加/删除部分、调整参数,所见即所得地创作3D内容。

复现评估

复现评估:\n\n开源情况:论文承诺提供开源实现,但需要检查代码仓库的完整性。\n\n数据可用性:PartRel3D数据集是新构建的,需要确认是否公开发布。该数据集包含约11K部分标注对象,跨越175个对象类别,超过90K个部分和300K规范化关系三元组。数据集验证采用两阶段协议:几何检查通过Open3D计算轴对齐边界框,对空间三元组应用谓词特定不等式过滤不一致三元组;人工审计均匀采样200个三元组,使用渲染的多视图图像和部分掩码手动验证正确性,重复20次得到稳定估计。空间三元组和功能三元组的平均正确率分别为92%和88%。\n\n算力需求:所有实验在4个NVIDIA L40 GPU上进行。训练使用AdamW,学习率 $1 \times 10^{-4}$,余弦衰减,梯度裁剪在1.0。这些是中等规模的资源需求,对于研究实验室是可及的。\n\n难度评估:总体难度中等。需要熟悉PyTorch、扩散模型、3D处理(Open3D、PointNet++)、以及注意力机制实现。依赖项包括3D VAE、图像VAE(如PixArt-α)、文本编码器(如Gemma 2)。数据处理需要解析三元组、渲染多视图图像、计算几何指标(CD、EMD、IoU)。评估需要多个基准数据集:Objaverse、ShapeNet、ABO、PartVerse、3D-Front。\n\n潜在障碍:部分数据集的获取和预处理可能耗时;VLM关系解析的配置需要额外步骤;3D渲染和几何计算的实现需要专业知识。但论文提供了详细的实现细节和评估协议,复现应该是可行的。