Tailor3D: Customized 3D Assets Editing and Generation with Dual-Side Images

1 Introduction

• Detailed customization and editing of 3D objects, such as adding patterns or changing styles, remain scarce despite advances in 3D generation.The introduction contrasts this gap with feed-forward methods that can create high-quality 3D objects in seconds.Source: §1 · p. 2
• Front and back views provide comprehensive object information with minimal overlap, simplifying independent edits compared with four-view editing.The paper notes that coordinating a localized change across four images can be difficult.Source: §1 · p. 2
• Tailor3D edits the front image, generates and edits the back image, then reconstructs a 3D object; each step takes seconds.The staged workflow gives users control over image textures before reconstruction.Source: §1 · p. 2
• The Dual-sided LRM handles inconsistencies between front and back views to improve reconstruction quality.It generates triplane features for each view and fuses them in 3D space using Viewpoint Cross-Attention.Source: §1 · pp. 2–3
• Tailor3D excels at local 3D generative fill, overall style transfer, and style fusion for objects.These applications are presented as evidence of the framework’s editing and customization capabilities.Source: §1 · p. 3

2 Related Work

• Multi-view diffusion evolved from Zero-1-to-3 [6] and SyncDreamer [7] toward fixed-view systems such as MVDream [13], ImageDream [14], and zero123++ [16].The passage describes geometric consistency as an ongoing challenge and notes that newer systems generate fewer overlapping canonical views.Source: §2 · p. 3
• LRM rapidly reconstructs 3D from single images, while Instant3D [11], LGM [12], and GRM [37] use multi-view generation or reconstruction approaches.Tailor3D reduces the input perspectives to front and back for extensive 3D editing, requiring lower consistency.Source: §2 · p. 3
• 3D editing has shifted from geometric and radiance-field methods toward editing 2D images and reconstructing 3D objects; MVEdit [56] takes 2–5 minutes.The passage contrasts this inference time with Tailor3D’s editing steps, which take seconds.Source: §2 · p. 3
• The Dual-sided LRM architecture converts front and back images into triplanes, then rotates and cross-attends their features to produce a 3D object.The figure caption describes this feature-fusion process as tailoring the two triplanes together.Source: Figure 2 · p. 4

3.1 Preliminaries

• LRM maps image features, initialized triplane embeddings, and camera extrinsics to three-plane features; multi-view diffusion generates a desired view from an image and camera parameters.The three planes are XY, YZ, and XZ; Tailor3D uses multi-view diffusion to generate the back image from the front.Source: §3.1 · pp. 4–5

3.2 The Pipeline of Tailor3D

• Tailor3D edits a front image, generates and edits its corresponding back image, then reconstructs a 3D object from the pair.The back image may instead be photographed or directly provided, and the original front and back images can also be reconstructed without editing.Source: §3.2 · p. 5

3.3 Dual-sided LRM: How to Accept Imperfect Consistent Views

• The method uses two views rather than four to reduce inconsistency pressure, then fuses flipped back triplanes with front-query/back-key-value windowed cross-attention.The attention window is 7; back-view scaling, rotation, and translation are each applied with a 10% probability during augmentation.Source: §3.3 · pp. 5–6
• LoRA replaces selected transformer projections to reduce memory use when adapting pre-trained LRM parameters.It replaces input/output projections for self-attention and q, k, v, and output projections for cross-attention.Source: §3.3 · p. 5
• Training supervises rendered front, back, and randomly selected side views using pixel-wise, perceptual, and total-variation losses.The total-variation loss is used to prevent image noise.Source: §3.3 · p. 6

4.1 Dataset: Gobjaverse-LVIS

• The 22K-object Gobjaverse-LVIS dataset supplies high-quality fine-tuning objects, with matched front and back views at identical elevation.Training also uses fixed front and back viewpoints and randomly selected side views for new-view synthesis.Source: §4.1 · p. 6

4.2 Implementation Details

• Implementation uses DINOv2 ViT-B/16 at 384×384, a 16-layer triplane transformer, and completes inference in less than 5 seconds.The LoRA rank is set to 4; training runs for 10 epochs on 8 A100 GPUs and takes about 6 hours.Source: §4.2 · pp. 6–7

4.3.1 Tailor3D Applications

• Tailor3D demonstrates local geometry and pattern filling through text or image prompts, including progressively adding a head, hands, and cloak to armor.Other examples add objects such as a mailbox, balloons, a flower bush, and a basketball hoop.Source: §4.3.1 · p. 7
• Style transfer and fusion accept image or text guidance and apply different styles to the front and back of objects while maintaining IP integrity.The demonstrations use Midjourney for 2D image generation and editing.Source: §4.3.1 · p. 7, Figure 4 · p. 8

4.3.2 Compare to Existing 3D Image-to-3D Generation Methods

• Qualitative comparisons show enhanced backside information with Dual-sided LRM; Wonder3D and TriplaneGaussian struggle with complex objects, while LGM exhibits ghosting and missing detail.The comparison uses a test set of 100 stable-diffusion-generated images; quantitative results and generation times are reported in Table 1, but their values are not supplied here.Source: §4.3.2 · p. 7

4.3.3 Ablation Study

• Viewpoint Cross-Attention outperforms convolutional or direct-addition fusion, rank 4 performs best among 2, 4, and 8, and front extrinsics alone yield accurate outcomes.The LRM structure accepts only front-facing camera parameters.Source: §4.3.3 · p. 7

5 Limitation and Conclusion

• Tailor3D outperforms other single-image-to-3D methods in the reported comparison, while qualitative results contrast its output with lower resolution or ghosting in alternatives.Wonder3D and TriplaneGaussian show lower resolution, while LGM often has ghosting with complex textures.Source: Table 1 · p. 9, Figure 5 · p. 9
• Reconstruction from only front and back views can be challenging for thick side profiles; meshes may remain low-resolution, and added geometric features may not substantially change them.The authors identify thicker side profiles and mesh quality as future-work targets.Source: §5 · p. 9

A Additional Introduction

• The supplementary material covers reconstruction background, LRM-family methods, training and implementation, Gobjaverse, additional experiments, and broader social impact.It is organized across Appendices B–E.Source: A Additional Introduction · p. 14

B Additional Related Work

• The related-work appendix groups reconstruction into single-view, multi-view, and normal-view categories, then discusses double-sided canonical reconstruction and LRM-family variants.The LRM family is presented as a universal reconstruction framework.Source: B Additional Related Work · p. 14

B.1 Single, Multi and Canonical-view Reconstruction

• Dual-Canonical-view Reconstruction uses front and back faces with fewer overlaps, making editing easier than editing four canonical faces simultaneously.Multi-view reconstruction requires per-object optimization, whereas canonical-view reconstruction uses a general framework.Source: §B.1 · p. 14

B.2 Introduction to LRM Family

• LRM directly reconstructs a mesh from one image; Tailor3D extends this framework by fusing front- and back-view 3D triplane features with Viewpoint Cross-Attention.Instant3D instead combines features from four views at the 2D image-feature level, while Tailor3D trains only Viewpoint Cross-Attention using pre-trained single-view weights.Source: §B.2 · pp. 14–16

C Additional Methodology

• Training uses the OpenLRM codebase, with parameter-count variations from the original LRM detailed in the appendix.The appendix also describes the experimental setup.Source: C Additional Methodology · p. 16
• The study uses 22K carefully rendered artificial images with high-quality textures and consistency, reporting satisfactory results with this relatively small dataset.Source: C Additional Methodology · p. 16
• The appendix adds Gobjaverse rendering details and further experiments, including additional Tailor3D examples and comparisons with multi-view reconstruction methods.It also discusses the process's broader social impact.Source: C Additional Methodology · p. 16

C.1 Training Settings

• The OpenLRM implementation uses 256 resolution and 40-dimensional triplane features, reduced from 512 and 80, while other listed model dimensions remain unchanged.Camera embeddings and the triplane transformer remain 1024-dimensional.Source: §C.1 · p. 16
• Training uses 96 rendering sample rays, a 3e −4 learning rate, 0.05 weight decay, and a cosine scheduler.Source: §C.1 · p. 16
• The model is trained for 20 epochs with batch size 16 across eight A100 GPUs.Source: §C.1 · p. 16

C.2 Dataset: Gobjaverse

• Gobjaverse renders two orbits around each object, with 24 higher-elevation views and 12 lower-elevation views, plus top and bottom views.The two orbits are shown in cyan and red, and top and bottom views in purple.Source: §C.2 · p. 16
• Training excludes top and bottom views, using opposite front-and-back directions along the x- and y-axes at the same z-axis elevation; the authors report higher consistency than diffusion-generated views.They attribute this consistency to higher data quality that facilitates feature fusion.Source: §C.2 · p. 16

C.3 Testset: 100 Images from Stable Diffusion

• The test set contains 100 Stable Diffusion-generated objects spanning animals, humans, plants, landscapes, and micro-scenes.Some additional qualitative examples may use Midjourney images.Source: §C.3 · p. 17
• With ground-truth back-side RGB images, Tailor3D shows more detailed texture and avoids holes compared with DreamComposer; the EscherNet comparison reports superior speed and quality.Source: Figure 10 · p. 19, Figure 11 · p. 19

• Appendix D.1 compares Tailor3D with more recent multi-view reconstruction techniques.Source: D Additional Experiments · p. 19
• Appendix D.2 presents additional Tailor3D examples similar to Figure 4.Source: D Additional Experiments · p. 19
• The added examples showcase the method’s object customization and editing abilities.Source: D Additional Experiments · p. 19

D.1 Comparison with More Multi-view Reconstructions

• 5 seconds versus several minutes, and superior 3D reconstruction results: Tailor3D is compared with optimization-based Dreamcomposer and EscherNet on GSO30 and Objaverse test data.The test set differs from the 100 SD test set used in the main paper.Source: §D.1 · p. 19
• With front and ground-truth back images, Tailor3D generates superior mesh results to DreamComposer, which tends to show more reconstruction defects.Source: §D.1 · p. 20
• Despite using 2 views to EscherNet’s 16, Tailor3D obtains better mesh results, although EscherNet generates more consistent images and has higher reconstruction quality.Source: §D.1 · p. 20

D.2 More Examples

• Additional qualitative examples show overall style transformations and localized edits through style transfer, style fusion, and 3D generative fill.Source: §D.2 · p. 20

• The authors describe Tailor3D as a direction for future 3D-generation research that delegates editing to 2D images and reconstruction to 3D.Source: E Broad Impacts · p. 20
• The authors cite qualitative results as evidence of practicality and report fine-grained operations with sub-second steps for industrial scenarios.Source: E Broad Impacts · p. 20
• The authors propose applications in animation and 3D game development, associating them with reduced production costs and broader creative content creation.Source: E Broad Impacts · p. 20