3D editing aims to modify an existing 3D asset according to user intent while preserving geometry and appearance outside the intended edit region. Existing methods often rely on task-specific training, source-3D inversion, or external 3D masks, and lack a unified framework for diverse editing conditions and tasks. We propose UMAS3D, a unified 3D editing framework that is training-free, inversion-free, and free from external 3D masks. UMAS3D converts text, bounding boxes, reference images, and their combinations into common target-view guidance, after which addition, deletion, modification, appearance, and pose editing share the same downstream 3D generation process. Our key insight is to couple online edit-region estimation with region-aware source preservation. Mask-Free Adaptive Region Modeling (MARM) initializes an Edit/Keep prior from source-target visual differences and progressively refines it as the target structure evolves, while Source-Guided Preservation Steering (SGPS) constrains the current Keep-region states to retain source geometry and appearance without suppressing target-driven changes. Experiments show that UMAS3D supports addition, deletion, modification, appearance, and pose editing while effectively preserving unedited source content. These advances make UMAS3D a unified framework for flexible and source-preserving 3D editing across diverse editing conditions.
Figures & tables
Figure 1: Overview of UMAS3D. A training-free framework that unifies multimodal conditions and five 3D editing tasks through adaptive region refinement and selective source preservation.
Figure 2: Core 3D editing pipeline of UMAS3D. MARM initializes and adaptively refines the Edit/Keep regions. SGPS preserves source-consistent content across structure and SLAT spaces.
Table 3
Unedited-region Preservation
Overall Target Quality
Method
Venue
Pres-IoU ↑
Pres-PSNR ↑
Pres-SSIM ↑
DINO-I ↑
CD-L1 ↓
F@1% ↑
NC ↑
EditP23
arXiv’25
0.1745
7.8001
0.4363
0.5598
0.0612
0.3684
0.6905
Easy3E
CVPR’26
0.5643
17.0366
0.7149
0.8568
0.0190
0.7719
0.8502
VoxHammer
3DV’26
0.5691
16.1752
0.6056
0.8468
0.0186
0.7598
0.8521
Nano3D
ICLR’26
0.5580
16.0651
0.6176
0.8554
0.0182
0.7696
0.8540
Vinedresser3D
CVPR’26
0.6030
19.8795
0.7913
0.8797
0.0158
0.8031
0.8712
Table 3: Quantitative comparison with existing 3D editing methods. Best results are highlighted in bold with a light red background.
Figure 3: Qualitative comparison with existing 3D editing methods.
Method
Instruction Alignment ↑
Visual Quality ↑
Source Preservation ↑
VoxHammer
7%
7%
8%
Nano3D
7%
9%
8%
Vinedresser3D
8%
11%
9%
3DEditFormer
21%
19%
14%
UMAS3D
57%
54%
61%
Table 4: Human evaluation results. Each value denotes the percentage of times a method is selected as the preferred result under the corresponding criterion.
Components
Unedited-region Preservation
Overall Target Quality
Model
Region Init.
S1 SGPS
S2 SGPS
Adapt.
Pres-IoU ↑
Pres-PSNR ↑
Pres-SSIM ↑
DINO-I ↑
CD-L1 ↓
F@1% ↑
NC ↑
A
–
–
–
–
0.1328
9.2586
0.4355
0.8644
0.0570
0.2861
0.6755
B
Ext.
✓
–
–
0.5437
16.0835
0.2619
0.8521
0.0162
0.7949
0.8633
C
Ext.
✓
✓
–
0.6438
17.2126
0.6770
0.8780
0.0153
0.8001
0.8800
D
Ext.
✓
✓
✓
0.6341
19.8069
0.6966
0.8798
0.0146
0.8048
0.8789
E
Mask-free
✓
✓
–
0.7094
18.6103
0.7737
0.8778
0.0152
0.8022
0.8959
Table 5: Quantitative ablation of UMAS3D components. Ext. : external 3D-mask initialization; Mask-free : MARM initialization without external 3D masks; S1/S2 SGPS : Stage 1/Stage 2 SGPS; Adapt. : adaptive Edit/Keep refinement. Best results are highlighted in bold with a light background.
Figure 4: Qualitative ablation of UMAS3D. Variants follow Table 5 . The first-row instruction is “Remove both bee wings.”, and the second-row instruction is “Replace the human arm structure with robotic mechanical arms.”
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
Input
Text
Bounding Box
Reference Image
T
[Original Instruction] The original editing instruction is retained verbatim without rewriting.
–
–
T+B
[Modality-Specific Text] Since the Bounding Box provides where the edit occurs, redundant spatial descriptions already conveyed by the box are removed, while preserving the editing operation and target content or appearance. The original editing intent remains unchanged.
[Four Source Views] + [Original Instruction] Select the source view that best exposes the edit-relevant region, identify the source-side edit anchor , and return a tight 2D box [x1,y1,x2,y2] normalized to [0,1000] . For Addition, the box localizes the source-side attachment or placement region rather than future target content absent from the source.
–
T+R
[Modality-Specific Text] Since the Reference Image provides what / appearance , the text retains the editing operation, spatial location, and necessary relation while removing target visual details already represented by the reference.
–
[Selected Source View] + [Original Instruction] Generate a standalone visual reference that represents only the target visual information required by the edit. For Addition, the reference depicts the object to be added; for Modification, the desired replacement or modified part; and for Appearance, the target color, material, texture, or pattern.
B+R
–
Uses the same source-side Bounding Box as T+B .
Uses the same Reference Image as T+R .
T+B+R
[Modality-Specific Text] The Bounding Box provides where , while the Reference Image provides what / appearance . The text therefore retains only the editing operation and any relation still required to specify the intended edit.
Uses the same source-side Bounding Box as T+B .
Uses the same Reference Image as T+R .
Appendix
Table 6: Construction of the five multimodal editing conditions used in our benchmark. Bracketed terms denote sample-specific content.
Figure 5: Examples of the constructed multimodal editing conditions. Each dashed box shows the Bounding Box visualization, the generated Reference Image, and the original textual editing instruction. Modality-specific text variants are constructed according to Table 6 .
Input
Gemini 3 Pro Image Input
Prompt Construction and Modality Roles
T
[Source View]
Text is the only explicit editing condition. The textual instruction defines the requested operation and target semantics, while the spatial location is inferred jointly from the instruction and the source image. The prompt explicitly prohibits introducing content that is not requested by the instruction.
T+B
[Source View] + [BBox Visualization]
Text defines what operation should be performed, whereas the Bounding Box provides only a soft cue for where the relevant source region is located. The box is not treated as a hard mask, exact boundary, crop, or target silhouette, and the actual edited area may extend slightly beyond or remain within the box when required for coherent geometry and appearance. The BBox itself must not appear in the generated image.
T+R
[Source View] + [Reference Image]
Text is authoritative for the editing operation, while the Reference Image provides the requested visual target information. Depending on the editing intent inferred from the available conditions, the reference may provide target identity, shape, parts, color, material, or texture. Its framing, pose, background, shadows, and unrelated content are not directly copied; instead, the relevant visual attributes are adapted to the source image.
No user-provided Text condition, original textual instruction, or edit-type label is supplied. The Bounding Box provides approximate where information, while the Reference Image provides what / appearance information. The editing operation is inferred from the source image, the localized source region, and the visual relationship between the source and the reference, and is represented internally by the unified editing description E .
All three modalities jointly constrain the edit. Text is authoritative for the requested operation / semantics , the Bounding Box provides soft where guidance, and the Reference Image provides the desired what / appearance . The spatial and visual conditions complement the textual instruction without overriding its editing intent.
Appendix
Table 7: Prompt construction for target-view generation under the five multimodal input combinations. The five settings share the same generation pipeline but differ in the roles assigned to Text, Bounding Box, and Reference Image.
Figure 6: Examples of target-view generation. Each dashed box groups a selected source view and its corresponding target view generated by Gemini 3 Pro Image. The examples cover diverse objects and editing operations.
Figure 7: Unified 3D editing across multimodal instruction combinations and editing tasks. UMAS3D converts heterogeneous multimodal conditions into target-view guidance and handles Addition, Deletion, Modification, Appearance and Pose through the same 3D editing pipeline.
Figure 8: Visualization of MARM initialization and adaptive region refinement. Each row shows the editing instruction, selected source view, generated target view, coarse 3D region prior, initial Edit/Keep partition M0 , and final partition MT after Stage 1 refinement. Differences between M0 and MT directly visualize the online reassignment of Edit/Keep regions within the single Stage 1 trajectory.
Component
Parameter
Value
TRELLIS
Stage 1 / Stage 2 steps
100/25
MARM
Adaptive regions K
8
Provisional macro regions
16
EMA coefficient λ
0.8
Initial warm-up
first 35 Stage 1 steps
Region update interval
every 5 steps
Appendix
Table 8: Main inference parameters of UMAS3D.
Figure 9: Additional qualitative 3D editing results of UMAS3D. Each dashed box presents four views of the source 3D and four corresponding views of the edited 3D, separated by a vertical dashed line.
Controllable local editing of 3D assets requires precise target localization and appropriate visual guidance. However, existing methods lack a simple yet accurate way to obtain 3D masks and struggle to achieve the desired edit while faithfully preserving the structure and appearance of non-target regions. To address these challenges, we present EditFlow3D, a training-free framework for local 3D editing. Given a source asset and an edit instruction, a VLM-driven workflow interprets the editing intent and automatically constructs a visual guidance image and a refined 3D editing mask, enabling localized editing in the native representation space of a pretrained 3D generative model. Specifically, mask-guided differential flow focuses the edit on the target region, while step-wise trajectory preservation maintains consistency between non-target regions and the source asset without directly replacing intermediate features. Since the existing Edit3D-Bench covers only a limited range of local editing categories, we further introduce EditFlow-Bench as a complementary benchmark encompassing a broader variety of structural and appearance edits, and evaluate EditFlow3D on both benchmarks. Quantitative results, qualitative comparisons, and a user study demonstrate that EditFlow3D achieves more accurate target-region editing and better preserves non-target regions than existing 3D editing methods.
Rui Nie, Chuang Wang, Haitao Zhou +4
Beihang University · Bambu Lab · Shanghai Jiao Tong University
Local editing of 3D objects remains a long-standing challenge. When interacting with 3D content, humans naturally tend to specify a coarse region of interest for modification rather than defining precise editing boundaries. However, previous methods rely on fully edited 2D images, precise 3D masks, or redundant pipelines, which present a gap. To bridge this gap, we propose EditVerse3D, a novel 3D editing framework that enables high-quality object editing under such coarse guidance. Our approach takes as input a 3D object to be edited, a coarse 3D bounding box indicating the target region, and a reference 2D image describing the desired modification. It produces a coherent, high-fidelity edited 3D object. To facilitate this editing, we introduce a novel region-aware adaptive loss that emphasizes hard-to-learn regions and balances the objective between target and preserved areas. Complementing our loss function, we enhance model robustness and generalization through targeted data augmentations, such as training with scaled 3D masks and filtering out unrealistic editing pairs. We construct a large-scale 3D editing dataset derived from parts information. Extensive experiments demonstrate that EditVerse3D achieves superior visual quality and quantitative performance compared to existing 3D editing approaches. Please visit our project page at https://editverse3d.github.io.
Youtan Yin, Yanning Zhou, Jiacheng Wei +6
College of Computing and Data Science, Nanyang Technological University · Tencent AIPD
Existing 3D scene editing methods typically rely on per-scene optimization over explicit 3D representations or cascaded edit-and-reconstruct pipelines, resulting in high test-time cost, limited 3D awareness, and structural inconsistencies. To couple appearance synthesis with geometry prediction, we adapt a pretrained unified RGB-geometry latent space to feed-forward scene editing. Given a source video and an edited reference image, JointEdit3D performs asymmetric latent inpainting: it observes only the edited RGB reference latent and jointly generates the remaining RGB latents and the entire geometry latent along the source trajectory. JointEdit3D introduces a dedicated SceneAnchor Branch to inject source-scene structure without forcing direct copying, and adopts edit/background-aware losses to balance edited-region fidelity with unedited-content preservation. To address the lack of paired resources for standardized 3D scene editing evaluation, we introduce SceneEdit3D-15K, a dataset with 15K paired editing samples and renderer-provided 3D annotations, together with SceneEdit3D-Bench, a curated 100-sample benchmark. Experiments show that JointEdit3D improves edited-region quality and 3D structural completeness over prior baselines while maintaining competitive background preservation.
Xinnan Zhu, Ruijie Xu, Jiayu Ying +4
East China Normal University · Shanghai Artificial Intelligence Laboratory · Fudan University +1