Recent diffusion-based methods have substantially advanced image editing, yet parametric 3D control at the part level remains challenging. Existing approaches either operate on whole objects with manual rigging, or revert to 2D handles at the part level. They also struggle to model appearance and texture under large transformations, leading to unrealistic outputs in geometry-intensive edits. To address these limitations, we introduce RADI3N, a framework for semantically-grounded part-aware object manipulation with parametric 3D control. Our CS-ARAP deformation module suppresses rotation-field discontinuities at part boundaries via neighborhood covariance smoothing, producing structurally consistent deformations under complex part-level edits. We further introduce an inversion-free selective texture transfer mechanism that suppresses high-frequency positional biases and scales up low-frequency semantic cues within cross-image attention. This eliminates strict spatial bias and supports semantic matching, allowing source appearance to follow the edited part onto its new geometry. Experiments demonstrate that RADI3N achieves state-of-the-art performance in geometric fidelity, viewpoint consistency, and controllability.
Table 1: Geometric editing capability matrix. Existing methods cover 2D editing and whole-object 3D parametric control. The combination of 3D parametric control with part-level granularity is uniquely addressed by RADI3N.
Figure 2: Overview of the geometric editing pipeline. Stage 1 (Coarse 3D Editing): The input image is lifted to a 3D mesh and deformed using CS-ARAP with covariance smoothing and a bi-Laplacian solve ( Lcoarse ) to obtain a smooth mesh. Intermediate renderings and source views condition the diffusion model via IP-Adapter for identity preservation. Stage 2 (Global 3D Refinement): The Stage 1 Jacobian field is refined using Lrefine , combining handle constraints with SDS gradients from the diffusion model, producing a semantically plausible mesh for final rendering.
Figure 3: Frequency-aware modulation of cross-image attention. For a query in the edited target (green marker), we visualize its attention over the source. Default RoPE attends to the part’s pre-edit location (a); this bias vanishes when high-frequency ( Shf ) chunks are removed (b) and attention shifts to semantically matching regions as low-frequency chunks ( Slf ) are amplified (c) Details in supp.
Method
Geometric Fidelity
Appearance & Identity
Consistency
WE ↓
D-RMSE ↓
MAAE ∘ ↓
Acc@5 ∘ ↑
CLIP-I ↑
FID ↓
SUBC ↑
DINOv2 ↓
BC ↑
GoodDrag [ 75 ]
0.086
2.77
26.84
0.084
0.894
167.90
0.797
1913.67
0.897
Image Sculpting [ 71 ]
0.082
2.75
19.83
0.293
0.897
160.46
0.808
1909.27
0.894
FlowDrag [ 28 ]
0.077
2.14
21.45
0.176
0.901
150.23
0.809
1901.10
0.933
FreeFine [ 77 ]
0.114
2.51
20.31
0.285
0.897
152.37
0.814
1901.23
0.821
Qwen-Img-Edit [ 69 ]
0.076
2.23
18.47
0.241
0.929
147.76
0.826
1896.16
0.952
Table 2: Quantitative comparison against state-of-the-art methods on GeoBench. RADI3N achieves the best geometric fidelity and appearance consistency across nearly all metrics, while maintaining competitive appearance and identity preservation.
Figure 4: Qualitative comparison with state-of-the-art methods. We compare RADI3N against other state-of-the-art methods across diverse object categories. Our method consistently produces geometrically faithful and photorealistic edits, while baselines exhibit artifacts, implausible geometry, or loss of scene context.
Figure 5: Qualitative results across editing operations. RADI3N supports diverse 6DOF control modes, including multi-part editing, part translation, and full-body rotation (F.B.R).
Figure 6: We compare ARAP and SR-ARAP with ours.
Table 3: Component Ablations and Efficiency Analysis.
Figure 7: Applications of RADI3N. We demonstrate the usability across several tasks. Details in supp.
Appendix figures & tables17 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 8: Extended qualitative comparisons across additional object categories. Each row shows the input, editing prompt, and outputs from all methods.
Figure 9: Additional results across editing operations: multi-part editing, part translation, and full-body rotation.
Figure 10: Part Segmentation Comparison.
Figure 11: Bounding box detection comparison across YoloWorld, OwlV2, and RADI3N. Tighter localizations lead directly to more accurate part segmentation in downstream stages.
Method
MAAE ∘ ↓
Acc@5 ∘ ↑
Image Sculpting
16.61
0.354
FlowDrag
23.93
0.141
Nano Banana 2
11.27
0.479
RADI3N
8.41
0.746
Appendix
Table 4: Part angular accuracy evaluated using Hunyuan3D on GeoBench for 3D reconstruction, independent of the SAM3D used in the editing pipeline.
Method
WE ↓
D-RMSE ↓
OA-Acc@5 ∘ ↑
FreeFine
0.126
3.11
0.573
Object3DiT
0.117
3.38
0.741
Qwen-Img-Edit
0.096
2.80
0.504
RADI3N
0.073
2.12
0.764
Appendix
Table 5: Full-body rotation evaluation. OA-Acc@5 ∘ (whether the desired angle is within the 5-degree range of the angle)is measured via Orient Anything.
Method
SUBC ↑
DINOv2 ↓
BC ↑
GoodDrag
0.804
1888.41
0.923
RADI3N
0.826
1863.17
0.941
Appendix
Table 6: Part-translation consistency evaluation.
Figure 12: User Study: Comparing our method with FreeFine, GoodDrag, FlowDrag, Qwen-Image-Edit, ImageSculpting
Method
Geometric Fidelity
Appearance & Identity
Consistency
WE ↓
D-RMSE ↓
MAAE ∘ ↓
Acc@5 ∘ ↑
CLIP-I ↑
FID ↓
SUBC ↑
DINOv2 ↓
BC ↑
GoodDrag
0.094
3.06
29.8
0.074
0.877
184.3
0.774
1933.4
0.873
FlowDrag
0.086
2.37
23.7
0.156
0.884
165.8
0.783
1923.1
0.910
FreeFine
0.125
2.76
22.5
0.252
0.879
168.2
0.789
1922.9
0.798
Image Sculpting
0.091
2.80
16.5
0.424
0.881
176.4
0.783
1929.4
0.881
Qwen-Img-Edit
0.084
2.47
20.5
0.213
0.908
162.9
0.799
1918.2
0.929
Appendix
Table 7: Quantitative comparison on DragBench .
Figure 13: 3D mask propagation via UV-space majority voting. Per-view 2D segmentations are lifted to consistent 3D part labels on the mesh surface.
Configuration
WE ↓
D-RMSE ↓
CLIP-I ↑
FID ↓
Vanilla SDS
0.146
2.12
0.898
152.25
SDS + IP-Adapter (ours)
0.067
1.91
0.924
148.47
Appendix
Table 8: Effect of IP-Adapter conditioning on SDS refinement.
Method
CLIP-I ↑
FID ↓
MasaCtrl
0.681
191.34
Cross-Image-Attn
0.803
199.82
RADI3N
0.924
148.47
Appendix
Table 9: Appearance transfer module comparison on GeoBench.
Model
Time (s) ↓
ULIP ↑
Uni3D ↑
SV3D
600
0.1124
0.3362
HY3D-2.1
200
0.1293
0.3546
TRELLIS
45
0.1473
0.3698
SAM3D
25
0.1488
0.3707
Appendix
Table 10: 3D Lifting Comparison. ULIP and Uni3D scores measure perceptual similarity between reconstructed geometry and the source image.
Figure 14: Qualitative comparison of appearance transfer methods. MasaCtrl and Cross-Image-Attn introduce color bleeding and structural artifacts that the full RADI3N module avoids.
Figure 15: Scene registration ablation. Each column removes one component of the blending pipeline, showing its contribution to the final composite quality.
Figure 16: Failure cases in mesh generation. Thin structures and heavily occluded surfaces produce incomplete or noisy meshes that limits edit quality.
Achieving precise, object-level control in image editing remains challenging: 2D methods lack 3D awareness and often yield ambiguous or implausible results, while existing 3D-aware approaches rely on heavy optimization or incomplete monocular reconstructions. We present ObjectMorpher, a unified, interactive framework that converts ambiguous 2D edits into geometry-grounded operations. ObjectMorpher lifts target instances with an image-to-3D generator into editable 3D Gaussian Splatting (3DGS), enabling fast, identity-preserving manipulation. Users drag control points; a graph-based non-rigid deformation with as-rigid-as-possible (ARAP) constraints ensures physically sensible shape and pose changes. A composite diffusion module harmonizes lighting, color, and boundaries for seamless reintegration. Across diverse categories, ObjectMorpher delivers fine-grained, photorealistic edits with superior controllability and efficiency, outperforming 2D drag and 3D-aware baselines on KID, LPIPS, SIFID, and user preference.
Single-image 3D generation with part-level structure remains challenging: learned priors struggle to cover the long tail of part geometries and maintain multi-view consistency, and existing systems provide limited support for precise, localized edits. We present PartRAG, a retrieval-augmented framework that integrates an external part database with a diffusion transformer to couple generation with an editable representation. To overcome the first challenge, we introduce a Hierarchical Contrastive Retrieval module that aligns dense image patches with 3D part latents at both part and object granularity, retrieving from a curated bank of 1,236 part-annotated assets to inject diverse, physically plausible exemplars into denoising. To overcome the second challenge, we add a masked, part-level editor that operates in a shared canonical space, enabling swaps, attribute refinements, and compositional updates without regenerating the whole object while preserving non-target parts and multi-view consistency. PartRAG achieves competitive results on Objaverse, ShapeNet, and ABO-reducing Chamfer Distance from 0.1726 to 0.1528 and raising F-Score from 0.7472 to 0.844 on Objaverse-with inference of 38s and interactive edits in 5-8s. Qualitatively, PartRAG produces sharper part boundaries, better thin-structure fidelity, and robust behavior on articulated objects. Code: https://github.com/AIGeeksGroup/PartRAG. Website: https://aigeeksgroup.github.io/PartRAG.
Precisely manipulating objects in a single photograph (translation, rotation, scaling) while obeying 3D physical constraints remains unsolved for diffusion-based editors. Current 2D methods lack spatial awareness and produce perspective violations. Forcing structural proxies into the latent space also disrupts variance homogeneity, and the resulting self-attention leakage leads to ghosting and background blur. The core difficulty is asymmetric: the relocated object must follow a rigid geometry, yet the uncovered background needs freedom to synthesize plausible content. We present GeoEdit, a training-free Lift-Manipulate-Render-Denoise pipeline that satisfies both constraints. We decouple scene and object in 3D, align them through point correspondence, and render a geometry-aligned proxy with a structural depth map. A Dual-Branch Denoising stage then refines this proxy: a video diffusion backbone preserves object identity, while 3D constraints are injected into the foreground within a narrow denoising window at matching noise variance (variance-homogeneous injection). The background denoises freely. Because the injected signal matches the native latent statistics, self-attention stays undisturbed. We also introduce GeoEditBench, a pose-aware benchmark covering object translation, object rotation, and camera movement with pose-aware evaluation metrics. Experiments confirm consistent gains in geometric accuracy, identity fidelity, and background quality. Our codes are available at https://github.com/Heey731/GeoEdit.
Yi He, Jiangming Wang, Xinyu Wang +5
Shenzhen International Graduate School, Tsinghua University · Sun Yat-sen University · Peking University +3