SpaceFlow

Locally Controllable 3D Generation

Training-free 3D generation with local geometric control
and part-specific text or image appearance conditioning.

Neil De La Fuente^{1 *}Joan Lafuente Baeza^{1 *}Mukhammadali Sayfiddinov^{1 *}Felicia Scharitzer^{1 *}

Marc Pollefeys¹ Ata Çelen¹ Sayan Deb Sarkar^{2 †}Elisabetta Fedele^{1 †}

¹ ETH Zürich² Stanford University* Equal contribution† Equal supervision

Paper & supplement Watch the teaser Interactive examples

TL;DR SpaceFlow generates 3D assets from geometric primitives, with local shape control and part-specific text or image appearance cues, without retraining.

Abstract

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, thereby limiting cross-part leakage. Regional geometry metrics demonstrate that SpaceFlow preserves the specified geometry in high-control regions and enables plausible shape variation in low-control areas. A user study further indicates that the resulting balance between geometric fidelity and generative freedom remains competitive in overall quality. When evaluating appearance on fixed geometry, text-conditioned routing achieves state-of-the-art prompt faithfulness and color/material accuracy. Qualitative results additionally show localized routing of image cues.

01 / OVERVIEW

Local geometric and
appearance control

SpaceFlow assigns geometric control strengths and appearance conditions to individual parts of a generated 3D asset.

Each input primitive represents an object part. Strong geometric guidance encourages adherence to the input shape, while weak guidance allows prompt-conditioned completion. During appearance generation, text or image cues are routed to the corresponding regions of the generated structure.

The pretrained generator remains frozen throughout both stages.

Geometric and appearance control

LOCAL GEOMETRY + LOCAL APPEARANCE

Paper overview: high-control chair legs and seat with a generated slat back; a gold trophy on black marble; a figurine with separate image references; and combined controls on a balloon, elephant, and lamp. — Selected examples from the paper. Orange primitives specify strong geometric control; gray primitives allow generative completion. Text and image references specify appearance locally. All six input/output pairs rotate in 3D; annotation arrows follow the corresponding input parts. The cactus is an image-conditioned example from the supplementary material.

INTERACTIVE EXAMPLES

Explore the generated assets

High control Preserve the input shapeLow control Allow generative completion

03 / METHOD

Structure generation and
appearance conditioning

SpaceFlow applies local guidance during sampling. Input primitives specify geometric constraints and associate text or image cues with object parts.

THE INPUT

The input consists of editable geometric parts, each with a local control level and an optional text or image appearance cue. Deformable superquadrics provide a compact representation for specifying these parts.

SpaceFlow pipeline from geometric inputs and local control through guided structure generation, part assignment, condition routing, and appearance generation. — The SpaceFlow pipeline from the paper. Select a stage above for a short explanation, or open the diagram to inspect the details.

Primitive-to-part appearance routing

Input primitives are matched to regions of the generated structure. Corresponding colors identify the same appearance condition; gray regions receive the global condition. These colors indicate cue assignments, not the final materials.

Pickup truck: input primitives colored by appearance-condition assignment. — **Pickup truck**
“red metal truck with a white cabin and blue glass”

Pickup truck: routing volume colored by appearance-condition assignment. — **Pickup truck**
“red metal truck with a white cabin and blue glass”

Toy elephant: input primitives colored by appearance-condition assignment. — **Toy elephant**
“blue wooden toy elephant with pink wooden ears”

Toy elephant: routing volume colored by appearance-condition assignment. — **Toy elephant**
“blue wooden toy elephant with pink wooden ears”

Enamel mug: input primitives colored by appearance-condition assignment. — **Enamel mug**
“white enamel mug with a blue handle”

Enamel mug: routing volume colored by appearance-condition assignment. — **Enamel mug**
“white enamel mug with a blue handle”

Selected examples from supplementary Figure S9. Routing is represented on a 32³ voxel grid using 30 PartField clusters.

Experimental results

SpaceFlow win rate

Evaluation criterion

vs. SpaceControl · τ = 375.9%

95% CI [65.7, 83.8]

vs. SpaceControl · τ = 1068.7%

95% CI [58.1, 77.6]

0%50% · equal preference100%

Overall preference balances fidelity to the input and realism. These are VLM-judge results over 83 assets, aggregated by majority vote across three passes; intervals are 95% Wilson confidence intervals.

05 / RESOURCES

Paper and resources

The paper, the workflow, and the generated assets.

↗Paper & supplementary materialMethod, experiments, and supplementary material ◈Interactive 3D galleryPaired inputs and results · rotate, compare, download GLBs ▷Video demonstrationGeometric and appearance-control workflow

Citation

@misc{delafuente2026spaceflow,
  title  = {SpaceFlow: Locally Controllable 3D Generation},
  author = {De La Fuente, Neil and Lafuente Baeza, Joan and
            Sayfiddinov, Mukhammadali and Scharitzer, Felicia and
            Pollefeys, Marc and Çelen, Ata and
            Deb Sarkar, Sayan and Fedele, Elisabetta},
  year   = {2026},
  note   = {Preprint}
}

Preprint citation. Publication metadata will be updated with the release.