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

Figure 1: Pipeline diagram for SemanticSlider3D with five labeled stages, using a dark brown upright piano and target attribute “very futuristic looking” as a running example. (a) Multi-View Rendering. The original dark brown upright piano is shown on the left. N camera views are rendered from uniformly distributed viewpoints and displayed in a 3-column grid, showing the piano from diverse angles including front-facing with visible keys, angled top-down views showing the lid, side profiles, and rear-facing views showing the flat back panel. Dotted lines indicate additional views not shown. (b) Generating Editing Prompt Pairs. GPT 5.2 generates two prompts shown in colored text boxes. The positive prompt in a blue box reads: “Make the piano ultra-sleek and futuristic, minimalist form, integrated LED accents.” The negative prompt in a yellow-orange box reads: “Make the piano classic and traditional, ornate wood finish, vintage design.” (c) Select Edit-Relevant Views. A vertical pink gradient bar labeled “Higher Relevance” indicates view ranking. Front-facing views showing the piano keys and body are ranked at the top, while less informative views such as rear-facing angles showing only the flat back panel are ranked lower. A bracket on the right identifies the M most relevant views selected for the next stage. (d) Generating Image Pairs. GPT Image 1.5 edits each selected view with both prompts. Two labeled columns show the results: the Positive column shows bright silvery futuristic piano versions with sleek angular forms; the Negative column shows dark brown traditional piano versions with ornate wooden appearance. Three visible pairs are shown with dotted lines indicating additional pairs below. (e) Editing via Generation. The TRELLIS pipeline combines multi-view conditions from stage (a) with the steering direction derived from the contrastive pairs in stage (d). Five 3D piano variations are shown at alpha values of −4.80, −1.25, 0, 2.13, and 5.00. At alpha −4.80 the piano is very dark brown with a heavy traditional appearance; at −1.25 it is a slightly lighter brown; at 0 it matches the original object; at 2.13 it has a lighter more modern form; at 5.00 it is a bright metallic blue with a sleek futuristic appearance. — Given a 3D object and a user-specified attribute, SemanticSlider3D renders multi-view images, generates contrastive editing prompt pairs, selects the most edit-relevant views, and converts the resulting contrastive image pairs into a steering direction — producing a coherent spectrum of 3D variations with no per-attribute training. Full-resolution PDF.

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.

Try the sliders

Each slider moves through nine variations produced by one semantic editing direction. The rotation control turns every object at once, so you can check that an edit holds up across views.

Rotate all 0°

An upright piano edited along a style attribute, from retro to futuristic; the slider sets the intensity and the rotation slider sets the viewing angle.

Piano · style

retro futuristic

A yellow car edited along a style attribute, from boxy to sporty; the slider sets the intensity and the rotation slider sets the viewing angle.

Car · style

boxy sporty

Book · condition

brand new well-worn

A standing figure edited along a local spatial attribute, from small eyes to big eyes; the slider sets the intensity and the rotation slider sets the viewing angle.

Figure · local spatial

small eyes big eyes

A tree edited along a temporal attribute, from bare to leafy; the slider sets the intensity and the rotation slider sets the viewing angle.

Tree · temporal

bare leafy

A quadruped animal edited along a fidelity attribute, from low-poly to photorealistic; the slider sets the intensity and the rotation slider sets the viewing angle.

Deer · fidelity

low-poly photorealistic

Video

Trouble with the embed? Watch on YouTube.

How it works

A user-specified attribute becomes a semantic editing direction in the latent space of TRELLIS — no per-attribute training.

Generating image conditions

Multi-view images are rendered from the 3D object, and an LLM generates contrastive editing prompts for the positive and negative extremes of the attribute. Edit-relevant views — those that actually carry visual information about the attribute — are ranked by the ImageReward score between the editing prompt and each camera view, verified by a VLM, then edited to produce contrastive image pairs.
Editing via generation

Steering directions are computed from the contrastive view pairs by averaging velocity differences across them. Combined with classifier-free guidance, the scaled steering is injected into both stages of the generation model to produce 3D variations at different attribute intensities. TRELLIS represents objects as Structured LATents (SLAT) — local latent vectors on a sparse 3D voxel grid — and its flow matching architecture makes this steering possible at inference time.
Mapping latent variations to the slider

Adaptive bounds are determined through quality checking, and perceptual distances are measured with LPIPS. Because generation does not guarantee consistent object orientation across α values, comparing key-frames by rendering order would inflate the measurement with orientation differences unrelated to the actual attribute change — so a minimum-cost bipartite matching pairs each original view with its closest-matching viewpoint in the variation. Variations are then mapped to slider positions by signed perceptual distance and refined iteratively with additional anchor points.

Evaluation

Technical validation

On a dataset of 50 3D object-attribute pairs spanning 46 unique 3D objects — covering local attributes (spatial, expressions) and overall characteristics (shape, style, material, detail, temporal) — 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.

Exploratory study

Six participants with prior 3D prototyping experience, ranging from one month to ten years, created objects and sliders across two tasks. The study demonstrates that SemanticSlider3D supported decision-making in 3D prototyping and was perceived as a valuable addition to existing workflows.

BibTeX

@article{wang2026semanticslider3d,
    title={SemanticSlider3D: Training-Free Continuous Semantic Editing for 3D Objects},
    author={Wang, Ru and Jain, Rahul and Niinuma, Koichiro and Gupta, Aakar},
    journal={arXiv preprint arXiv:2608.18560},
    year={2026}
}

Acknowledgments

SemanticSlider3D builds on TRELLIS for structured-latent 3D generation. The playground additionally uses SAM 3D Objects for single-image reconstruction.