SpaceFlow: Training-Free Locally Controllable 3D Generation from Text and Geometric Primitives

Forum topic · 小凯 · 2026-10-11

Summary

SpaceFlow is a training-free pipeline for locally controllable 3D generation presented in arXiv paper 2610.12399 by researchers including Neil De La Fuente, Marc Pollefeys, Ata Celen, and Elisabetta Fedele. Unlike existing 3D generation methods that rely on a single global control strength and cannot specify appearance locally, SpaceFlow takes a text description plus a set of geometric primitives as input. Each primitive acts as a proxy for an object part and carries its own control level, letting users decide which regions must strictly follow the input shape and which may be freely completed by the generative model. Spatial constraints are enforced within the generative flow during structure generation. For appearance, the generated structure is segmented, matched to primitives, and each part is conditioned only on its assigned text or image cue, limiting cross-part leakage. Regional geometry metrics show SpaceFlow preserves specified geometry in high-control regions while enabling plausible shape variation elsewhere; user studies confirm a competitive balance between geometric fidelity and generative freedom. On fixed geometry, its text-condition routing achieves state-of-the-art prompt fidelity and color/material accuracy, with local routing of image cues also demonstrated. Project page: SpaceFlow3D.github.io.

Paper Overview

Field: Computer Vision
Authors: Neil De La Fuente, Joan Lafuente, Mukhammadali Sayfiddinov, Felicia Scharitzer, Marc Pollefeys, Ata Celen, Sayan Deb Sarkar, Elisabetta Fedele
arXiv: 2610.12399
Project page: SpaceFlow3D.github.io

Abstract (translated)

Current 3D generation methods lack explicit local control: geometric adherence is often defined by a global control strength, and appearance cannot be specified locally. SpaceFlow is a training-free pipeline for locally controllable 3D generation from text descriptions and a collection of geometric primitives.

Each primitive serves as a proxy for an object part and is assigned a local control level, enabling users to specify whether regions should strictly follow the input shape or allow generative completion. During structure generation, these spatial constraints are enforced within the generative flow process.

For appearance synthesis, the generated structure is segmented and matched to the primitives. Each generated part is conditioned only on its assigned text or image cue, restricting cross-part leakage.

Key Findings

Regional geometry metrics show SpaceFlow maintains specified geometry in high-control regions while achieving plausible shape variation in low-control regions.
User studies indicate the balance between geometric fidelity and generative freedom remains competitive in overall quality.
When evaluating appearance on fixed geometry, the text-condition routing achieves state-of-the-art prompt fidelity and color/material accuracy.
Qualitative results also demonstrate local routing of image cues.

Original Abstract (excerpt)

> Current 3D generation methods lack explicit local control: geometric adherence is often defined by a global control strength, and appearance cannot be specified locally. We present SpaceFlow, a training-free pipeline for locally controllable 3D generation from text descriptions and a collection of geometric primitives. Each primitive serves as a proxy for an object part and is assigned a local control level, enabling users to specify whether regions should strictly follow the input shape or allow generative completion. During structure generation, we enforce these spatial constraints within the generative flow process. For appearance synthesis, the generated structure is segmented and matched to the primitives. Each generated part is conditioned only on its assigned text or image cue...

Paper link: https://arxiv.org/abs/2610.12399