Source-linked AI summary
NANO3D: A Training-Free Approach for Efficient 3D Editing Without Masks
Junliang Ye, Shenghao Xie, Ruowen Zhao, Zhengyi Wang, Hongyu Yan, Wenqiang Zu, Lei Ma, Jun Zhu
TL;DR
3D editing lacks efficient, consistent training-free methods that preserve unedited regions. Nano3D integrates FlowEdit with TRELLIS and region-aware merging, achieving strong editing performance while enabling Nano3D-Edit-100k, a dataset of over 100,000 samples.
Problem
Existing 3D editing methods struggle with consistency, efficiency, and preservation of unedited regions, limiting practical asset refinement.
Method
Nano3D integrates FlowEdit into TRELLIS and uses Voxel/Slat-Merge for localized mask-free editing with source-region preservation.
Results
Nano3D achieves state-of-the-art performance across diverse 3D editing tasks and outperforms existing baselines in reported evaluations.
Takeaways & Limitations
Nano3D-Edit-100k provides over 100,000 high-quality 3D editing samples for future feedforward 3D editing research.
Takeaways & Limitations
Nano3D supports only localized edits, while TRELLIS reconstruction loss and generative capacity constrain overall performance.
Abstract
from arXiv · showhide
3D object editing is essential for interactive content creation in gaming, animation, and robotics, yet current approaches remain inefficient, inconsistent, and often fail to preserve unedited regions. Most methods rely on editing multi-view renderings followed by reconstruction, which introduces artifacts and limits practicality. To address these challenges, we propose Nano3D, a training-free framework for precise and coherent 3D object editing without masks. Nano3D integrates FlowEdit into TRELLIS to perform localized edits guided by front-view renderings, and further introduces region-aware merging strategies, Voxel/Slat-Merge, which adaptively preserve structural fidelity by ensuring consistency between edited and unedited areas. Experiments demonstrate that Nano3D achieves superior 3D consistency and visual quality compared with existing methods. Based on this framework, we construct the first large-scale 3D editing datasets Nano3D-Edit-100k, which contains over 100,000 high-quality 3D editing pairs. This work addresses long-standing challenges in both algorithm design and data availability, significantly improving the generality and reliability of 3D editing, and laying the groundwork for the development of feed-forward 3D editing models. Project Page:https://jamesyjl.github.io/Nano3D
1 INTRODUCTION
3D editing remains difficult because existing approaches are inefficient, inconsistent, and often fail to preserve unedited regions. Nano3D addresses this gap with training-free FlowEdit-based editing, region-aware merging, and a large paired dataset.
- Existing 3D editing methods struggle with cross-view and attribute consistency while requiring time-consuming optimization.
- Nano3D introduces FlowEdit into TRELLIS to enable efficient training-free 3D editing using pretrained generative priors.
- Voxel/Slat-Merge adaptively preserves source consistency by identifying edited regions and retaining unedited structure.
- Nano3D-Edit-100k contains over 100,000 high-quality 3D editing samples for future research and model development.
- Nano3D provides user-friendly localized editing without manual masks and supports state-of-the-art editing performance.
2 RELATED WORK
Prior 3D editing methods commonly optimize 3D representations from 2D guidance or edit multiple views before reconstruction. These strategies face substantial spatial-consistency challenges compared with 2D editing.
- 2D editing methods include inversion-based editing, finetuning on paired data, attention manipulation, adapters, and rectified flows.
- 3D generation research spans GAN-based and diffusion-based paradigms for modeling or denoising 3D representations.
- Maintaining spatial consistency is substantially harder in 3D than in 2D image editing.
- SDS-based methods optimize 3D representations using gradients from pretrained 2D diffusion models.
- Other approaches edit multi-view images and reconstruct them with large reconstruction models.
3 PRELIMINARY
FlowEdit performs inversion-free, optimization-free editing by constructing a latent trajectory between source and target conditions. TRELLIS represents 3D assets with structured latents and separates structure prediction from later latent processing.
- FlowEdit constructs an ODE trajectory from source to target prompts, using combined velocity fields to shorten editing paths and preserve structure.
- FlowEdit is inversion-free, optimization-free, and model-agnostic for text-guided editing with text-to-image flow models.
- TRELLIS represents assets with activated voxels carrying local latent vectors that jointly encode geometry and appearance.
- TRELLIS uses successive stages in which a 64^3 voxel grid estimates occupancy to produce a sparse structural representation.
4 METHOD
Nano3D combines FlowEdit with TRELLIS’s geometry and appearance stages, then merges edited and source regions to preserve consistency. The same framework supports automated construction of filtered 3D editing data.
- 4 METHOD: Nano3D voxelizes and encodes the input, applies FlowEdit-guided geometry editing, generates structured latents for appearance editing, and decodes the merged result.
- 4 METHOD: FlowEdit uses the input object’s front view as the source condition and an edited front view as the target to generate edited voxels.
- 4.3 VOXEL/SLAT-MERGE: Voxel-Merge computes voxel differences, selects significant connected components, and transfers edited regions while keeping non-edited regions unchanged.
- 4.3 VOXEL/SLAT-MERGE: SLat-Merge applies the previously defined mask to merge edited latent features with those encoded from the original asset.
- 4.4 NANO3D: The dataset pipeline samples frontal views, generates instructions with Qwen2.5-VL, reconstructs assets with TRELLIS, synthesizes edited images, and filters Nano3D outputs.
- 4.4 NANO3D: TRELLIS-reconstructed meshes are used because extracting structured latents from original meshes requires rendering about 150 views and still incurs VAE divergence.
5 EVALUATION
Nano3D achieves strong localized, view-consistent 3D edits while preserving unedited structure, outperforming baselines quantitatively and in user preference. Ablations show that Voxel-Merge and Slat-Merge provide complementary geometry and appearance consistency.
- 5.2 MAIN RESULT: Qualitatively, Nano3D follows editing instructions and preserves source structure across views, whereas baselines exhibit distortions, artifacts, blur, or semantic misalignment.Tailor3D, Vox-E, and TRELLIS each show distinct failure modes in the comparison.
- 5.2 MAIN RESULT: Nano3D achieves the lowest CD and FID and the highest DINO-I score among all baselines.These metrics assess structure preservation, generation quality, and semantic alignment, respectively.
- 5.2 MAIN RESULT: Nano3D received the highest user preference across Prompt Alignment, Visual Quality, and Shape Preservation in a 50-participant study.Tailor3D and Vox-E results were omitted because participants mainly favored TRELLIS and Nano3D.
- 5.3 ABLATION STUDY: Voxel-Merge restores geometry and cross-view consistency, while adding Slat-Merge further improves local visual quality and appearance consistency.FlowEdit alone produces geometry misalignments, deformations, and appearance inconsistencies.
- 5.3 ABLATION STUDY: For the voxel-merge threshold, τ = 100 most accurately matches editing regions, whereas lower thresholds include irrelevant non-editing areas.The comparison evaluates τ = 100, τ = 50, and τ = 30 for wing removal.
- 5.4 COMPARISON WITH EDITING METHODS REQUIRING MASKED INPUT: Nano3D preserves non-edited regions better than VoxHammer under identical mask-guided settings while also supporting editing without user-provided masks.The comparison uses identical bounding boxes, instructions, and target outputs.
6 CONCLUSION
Nano3D is a training-free framework for localized 3D object editing that combines FlowEdit with region-aware Voxel/Slat-Merge strategies. The work also introduces Nano3D-Edit-100k and identifies limitations involving edit scope, VAE reconstruction loss, and TRELLIS capacity.
- 6 CONCLUSION: Nano3D integrates FlowEdit into TRELLIS with Voxel/Slat-Merge to produce geometrically consistent and semantically faithful localized edits.The framework supports object removal, addition, and replacement.
- 6 CONCLUSION: Nano3D-Edit-100k is presented as the first large-scale dataset tailored for 3D editing and supports future feedforward DiT-based editing research.The dataset is constructed from the Nano3D framework.
- 6 CONCLUSION: Nano3D supports only localized edits, incurs reconstruction loss from the TRELLIS VAE, and remains constrained by TRELLIS’s generative capacity.The authors identify these limitations as directions for future research.
A.1 MORE VISUALIZATION RESULTS
Nano3D preserves geometric and textural consistency across additional addition, removal, and replacement edits.
- Nano3D effectively preserves geometric and textural consistency before and after addition, removal, and replacement edits.
A.2 CHOICE OF 3D REPRESENTATION: VOXEL VS. VECSET
FlowEdit is unstable in Hunyuan2.1, with outcomes ranging from fragmented shapes to unchanged source-like meshes depending on hyperparameters.
- FlowEdit fails reliably in Hunyuan2.1: aggressive hyperparameters cause fragmented or mud-like assets, while conservative settings ignore the target condition.The authors attribute the difference to Hunyuan2.1’s global vecset representation versus TRELLIS’s local voxel representation.
A.3 EFFECT OF IMAGE CONSISTENCY IN FLOWEDIT EDITING
Inconsistent pre- and post-edit renderings in 3D-Alpaca undermine localized FlowEdit editing and produce unstable outputs.
- Inconsistent 2D conditions cause FlowEdit to fail localized editing, producing either source-like outputs or entirely inconsistent results.Large n max preserves the source and ignores the target, whereas small n max disregards the source condition.
- 3D-Alpaca’s poor image consistency includes cabinet displacement and character-scale changes that lead to mismatched 3D assets.
A.4 THE PROMPT USED TO GENERATE EDITING INSTRUCTION FROM THE RENDERING
The appendix presents the instruction-generation prompt and category distribution alongside additional editing visualizations.
- A strict VLM prompt template constrains generated editing instructions to Nano3D’s supported capabilities.
- Additional visualizations cover addition, removal, and replacement operations.