SemanticSlider3D: Training-Free Continuous Semantic Editing for 3D Objects

Ru Wang; Rahul Jain; Koichiro Niinuma; Aakar Gupta

cs.HCAug 19, 2026

SemanticSlider3D: Training-Free Continuous Semantic Editing for 3D Objects

Authors: Ru Wang, Rahul Jain, Koichiro Niinuma, Aakar Gupta

Organizations: Department of Computer Sciences University of Wisconsin–Madison Madison, Wisconsin, USA · Fujitsu Research of America Pittsburgh, Pennsylvania, USA · Department of Electrical and Computer Engineering Purdue University West Lafayette, Indiana, USA · Fujitsu Research of America Redmond, Washington, USA

Abstract

Fine-grained control over continuous semantic attributes of 3D objects is essential for 3D content creation, but is not well supported by conventional 3D modeling workflows or prompt-based interaction with existing generative AI tools. While slider-based methods have proven effective for fine-grained semantic control in 2D image generation, no equivalent approach exists for 3D. Extending these 2D methods to 3D is non-trivial due to challenges unique to 3D, including geometric integrity and cross-view coherence. We present SemanticSlider3D, a technique for continuous semantic attribute editing of 3D objects that requires no per-attribute training. Given a user-specified attribute, our pipeline constructs a semantic editing direction in the latent space of a state-of-the-art 3D generation model, presenting a diverse and coherent spectrum of 3D variations. A technical validation on a dataset of 50 3D object-attribute pairs shows our method was preferred by all five human assessors across variation range, consistency, 3D object quality, and attribute disentanglement, over a baseline combining a 2D slider with an image-to-3D model. An exploratory study with six participants demonstrates that SemanticSlider3D supported decision-making in 3D prototyping and was perceived as a valuable addition to existing workflows.

Figures & tables

Figure 1. The pipeline of generating image conditions and editing through generation process. We use a piano as an example, with "very futuristic looking" being the target attribute specified by the user. (a) Rendering $N$ camera views from the original 3D object; (b) generating contrastive editing prompts for both positive direction and negative direction based on user-specified attribute; (c) ranking and selecting camera views relevant to the editing prompts; (d) generating contrastive image pairs using edit-relevant views and editing prompt pairs for image editing; (e) Editing 3D objects along the semantic axis using TRELLIS with multi-view conditions and steering direction defined by contrastive image pairs.

Figure 2. The interactive playground interface used in our study. Users generate 3D objects via image-to-3D (a: upload and segment a reference image (a1) to produce a 3D object (a2)) or text-to-3D (b: type a description (b1) to generate a 3D object (b2)). The Slider Panel (c) allows users to specify an object and semantic attributes, create a slider, and add new anchor points at desired percentages; as the user moves the slider handle on created sliders, the corresponding variation is displayed in-place in the viewport (f). The Sidebar (d) displays thumbnails of all objects in the scene. The Transform Panel (e) provides controls for positioning, rotating, and scaling selected objects.

Figure 3. P3 created a (a) “futuristic style” slider for a floor lamb generated from an image, and chose the -98% variation to further refine with a (b) “umbrella-shaped hood” slider and chose 74% variation to further edit with a (c) “pattern abstractness” slider to get a desired pattern. Green % = input object’s slider position (same in later figures).

Figure 4. P6 created an (a) “abstractness” slider on an object generated from an image (a man hugging a ball) sourced online, with the goal of designing a sculpture with the same gesture. She then chose a variation at 42% and further explored material use with a (b) “hard vs. soft” slider where she found a good design at -100%.

Figure 5. (a) P1: no entrance for bartender at 96% on a “clutteredness” slider for a bar counter; (b) P5: skinny deer at -84% on a “strong” slider.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Stage	Time (s)
Multi-view rendering	$19.0\pm 2.3$
Prompt pairs generation	$2.1\pm 0.7$
Best-view selection + filtering	$21.2\pm 14.4$
Image pairs generation	$43.6\pm 16.3$
Image condition generation	$22.0\pm 1.2$
3D generation (per variation)	$65.3\pm 16.3$

Appendix

Table A.1. Per-stage time breakdown of our 3D editing pipeline (

N

=40,

M

=10). Values are reported as mean

\pm

standard deviation in seconds. Multi-view rendering through image condition generation are one-time costs for each slider. 3D generation repeats once per sampled variation, while GLB exporting repeats once per exported slider position.

ID	Age/Gender	Experience	3D Prototyping Tools
P1	25/M	1.5 yr	Unity, Unreal, Blender
P2	26/F	2 yr	Unity, Blender
P3	28/M	10 yr	Fusion, SolidWorks, Blender, Unity
P4	22/F	1 mo	SolidWorks
P5	24/F	6 yr	Fusion, SolidWorks
P6	31/F	10 yr	Fusion, SolidWorks

Appendix

Table A.2. Demographic information and 3D prototyping experience of participants.

Figure A.1. The process of mapping slider positions to $\alpha$ values. In this example, the 3D object is an airplane, and the target attribute is "rustiness". (a) Sampling $K=9$ evenly distributed initial values of $\alpha$ ; (b) building the mapping between perceptual distance and slider positions, and using initial anchors for future refinement; (c) interpolating additional anchor points on the slider based on existing anchors to refine the slider mapping; (d) comparison result showing slider variation is smoother and more consistent after slider mapping.

Figure A.2. The evaluation interface for the technical validation, showing the reference 3D object and target attribute (top left), two sequences of nine 3D variations generated by SemanticSlider3D and the baseline (bottom), and evaluation questions (top right). The mapping between methods and Slider A/B labels is randomized per object-attribute pair; in this example, Slider A corresponds to SemanticSlider3D and Slider B to the baseline. The target attribute is “clothing thickness (thin to thick).”

Figure A.3. Assessors’ response to the Likert scale questions.

Figure A.4. Mean assessor ratings across attribute categories.

3D Asset Editing

See all 18 papers

ES3D: Embedding Semantics into 3D Space for Component-Aware Editing
Aug 16, 2026Xuancheng Jin, Rengan Xie, Jiayuan Lu +53D Asset Editing 3D Shape Editing
State Key Laboratory of CAD&CG, Zhejiang University · NetEase Fuxi AI Lab
Vinedresser3D: Agentic Text-guided 3D Editing
Feb 23, 2026Yankuan Chi, Xiang Li, Zixuan Huang +1Text-Guided 3D Scene Editing 3D Asset Editing
The Hong Kong University of Science and Technology · University of Illinois Urbana-Champaign
EditVerse3D: High-Quality 3D Object Editing with Region-Aware Learning
Jul 8, 2026Youtan Yin, Yanning Zhou, Jiacheng Wei +63D Shape Editing 3D Scene Editing
College of Computing and Data Science, Nanyang Technological University · Tencent AIPD

Aug 16, 2026J/K move · Enter open · S save

ES3D: Embedding Semantics into 3D Space for Component-Aware Editing

Xuancheng Jin, Rengan Xie, Jiayuan Lu, Wenting Zheng+4

State Key Laboratory of CAD&CG, Zhejiang University · NetEase Fuxi AI Lab

3D Asset Editing 3D Shape Editing 3D Scene Representation

Open paper AskPDF arXiv

Existing 3D editing methods have made notable progress in controllability, yet they remain limited in several important ways. Most approaches rely on text-driven editing, which struggles to express fine-grained visual changes intended by the user. Moreover, many methods require manually supplied 3D masks or introduce unintended changes to regions that should remain untouched. These limitations largely arise from the absence of fine-grained semantic understanding, making it difficult for existing models to retrieve or modify specific 3D components. We introduce ES3D, a framework that embeds semantics directly into 3D space, enabling component-aware retrieval and editing of a 3D asset conditioned on multiple local reference images and optional text queries. We first construct a 3D semantic embedding by projecting multi-view semantic features into the voxelized space of the asset. We then perform 3D component retrieval by computing feature similarity between the 3D semantic embedding and the semantic embeddings of image or text queries. For editing, we employ a pretrained 3D generative model with an inpainting mechanism to modify the retrieved components guided by user-provided images while preserving the rest of the asset. Overall, ES3D is a 3D editing framework that retrieves editable regions based on semantic cues and uses multiple images as conditions. Extensive experiments demonstrate that ES3D produces geometrically consistent and semantically coherent edits, enabling robust image-based and text-assisted control for 3D editing.

Abstract

3D Asset Editing

Explore similar work

ES3D: Embedding Semantics into 3D Space for Component-Aware Editing

Vinedresser3D: Agentic Text-guided 3D Editing

EditVerse3D: High-Quality 3D Object Editing with Region-Aware Learning