NANO3D: A Training-Free Approach for Efficient 3D Editing Without Masks

TL;DR

Nano3D tackles the challenge of training-free, localized 3D object editing without masks by embedding FlowEdit into the TRELLIS framework and introducing region-aware Voxel/Slat-Merge merging to preserve unedited regions. The approach decouples geometry and appearance editing across two stages, enabling coherent, view-consistent edits guided by front-view renderings, and it yields a high-quality large-scale editing dataset, Nano3D-Edit-100k. Empirical results show state-of-the-art 3D consistency and visual fidelity against baselines, supported by qualitative, quantitative, and user-study evidence. This work advances practical, data-efficient 3D editing and lays the groundwork for feed-forward 3D editing models trained on real editing data.

Abstract

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

Figures (10)

• Figure 1: Highly-consistent 3D objects edited by Nano3D. Our framework supports a range of training-free and part-level tasks especially on shape, including removal, addition, and replacement, while only requiring users to provide source 3D objects and instructions, without any mask.
• Figure 2: The Nano3D pipeline. The original 3D object is voxelized and encoded into sparse structure and structured latent respectively. Stage 1 modifies geometry via Flow Transformer with FlowEdit, guided by Nano Banana–edited images. Stage 2 generates structured latents with Sparse Flow Transformer, supporting TRELLIS-inherent appearance editing. Voxel/Slat-Merge further ensures consistency across both stages before decoding the final 3D object.
• Figure 3: Data Construction Pipeline. The figure shows our pipeline. We first sample images from the dataset and prompt Qwen2.5-VL to generate editing instructions by completing templates. Trellis then generates 3D meshes from the images. Finally, the image, instruction, and mesh are fed into Nano3D, and the resulting 3D assets are filtered for quality.
• Figure 4: Qualitative results. We present three edit types—object removal, addition, and replacement. In each case, Nano3D confines changes to the target region (red dashed circles) and produces view-consistent edits, while leaving the rest of the scene unchanged. Geometry stays sharp and textures remain faithful in unedited areas, with no noticeable artifacts.
• Figure 5: Qualitative comparison. Our method achieves the best editing performance with faithful instruction semantic alignment and perfect original structure consistency across multi-view images.