SpaceFlow: Locally Controllable 3D Generation

arXiv:2610.12399 · cs.CV, cs.AI, cs.GR · Submitted 2026-10-08 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "SpaceFlow: Locally Controllable 3D Generation".

Jane: SpaceFlow presents a training-free pipeline for locally controllable 3D generation from text descriptions and geometric primitives,

Tom: First, who's behind it and why it matters.

Paper summary: Tom: So to wrap up this discussion on SpaceFlow: Locally Controllable three dee Generation, we're talking about a training-free pipeline that lets users specify whether different regions of a three dee asset should follow an input shape strictly or allow the AI to generate something new <ref:2610.12399#pg1,SpaceFlow: Locally Controllable 3D Generation>.

Jane: The authors, including Neil De La Fuente and others at ETH Zürich and Stanford University, present this unified framework for local geometric control and local appearance conditioning.

Lu: What this means in practice is that creators get to be much more interactive with the generative process by giving explicit instructions on which parts need fidelity versus where they can use the generative prior.

Meng: It feels like it's moving away from just getting one good overall render toward building assets piece by piece with precise local guidance.

Tom: It’s about giving the user agency over how much they want to trust the AI's creative freedom versus how much they need it to adhere to their initial design constraints.

Jane: The paper shows that this approach allows for high fidelity in constrained areas while still letting the generative prior handle the rest of the details based on text cues.

Tom: It’s a lot of control, but it keeps the process training-free, which is a big deal for making these tools accessible in professional workflows right now.

Lu: The future work mentioned suggests they are looking at extending this to even more complex control scenarios beyond just geometry and appearance on primitives.

Jane: So while the paper shows a solid way to achieve local control, the challenge ahead is handling those moments when the text prompt and the required geometry really start fighting each other.

Conclusion: Tom: So, we’ve been looking at SpaceFlow today. It’s about taking three dee generation from a global level down to a local level where you can tell the AI exactly where it needs to stick to your shape.

Jane: Right. The authors are Neil De La Fuente and others from ETH Zürich and Stanford University. They built this training-free pipeline for locally controllable three dee generation using text descriptions and geometric primitives like superquadrics.

Lu: What’s really interesting is how they handle appearance separately from structure. Instead of the whole thing being guided by one prompt, each generated part gets its own local cue—either a text string or an image—so you can condition it on different things.

Meng: From an engineering standpoint, the way they do that routing for appearance sounds clever because it keeps cross-part leakage low. It means when you’re coloring a specific piece, it only cares about its assigned part cue and not what’s happening on the adjacent parts.

Lalam: And Lalam sees this as a huge cultural shift for how we design things. If creators can build assets piece by piece with local fidelity, it opens up possibilities for modular content creation in game development or even virtual world building where you need fine control over specific components.

Tom: Exactly. The main thing is that they show you the balance between geometric accuracy and creative freedom remains really competitive when you look at user studies. It’s not just about making one perfect whole model anymore; it’s about building a model piece by piece with intent.

Jane: And the results on fixed geometry are telling, showing that text-conditioned routing achieves really high prompt faithfulness for color and material accuracy on those parts they constrain tightly.

Lu: The limitation they point out is when the text prompt fights the input geometry. If you ask for something impossible given your starting shape, it struggles to reconcile those two signals. It’s a gap in their joint prior right now.

Meng: That makes sense practically; if you tell the AI to put a tiny wire inside a solid block, it won't know how to handle that conflict without extra guidance or a fix.

Tom: So, SpaceFlow gives you that explicit control mechanism—the local control levels—to manage that tension between following your blueprint and letting the generative prior fill in the gaps.

Jane: It moves us toward a generation process where you don't just get a result; you have specific knobs to turn for every part of that result. (Music swells slightly)

Neil De La Fuente, Joan Lafuente, Mukhammadali Sayfiddinov, Felicia Scharitzer, Marc Pollefeys, Ata Çelen

ETH Zürich · Stanford University · Microsoft

cs.CV, cs.AI, cs.GR

Submitted: 2026-10-08

Updated: 2026-10-08

Code: https://github.com/GabrielTavernini/superflex

Project page: https://spaceflow3d.github.io

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 75/100

The gist: SpaceFlow presents a training-free pipeline for locally controllable 3D generation from text descriptions and geometric primitives, enabling users to specify whether regions should strictly follow an

Key concepts

Local Control Levels
Each geometric part is given a control level, ranging from strict adherence to complete freedom. Higher levels enforce the input shape closely, while lower levels permit the generative model more flexibility. This allows users to specify exactly which parts should be fixed and which can be creatively generated.
Local Appearance Routing
This mechanism ensures that appearance cues (like colors or materials) only affect their designated 3D part. It maps every voxel in the generated structure to a specific input cue, preventing appearance information from leaking across different object sections during synthesis.
Geometric Primitives
These are simple building blocks, like basic shapes, used as proxies for object parts. They are assigned control levels to define the geometric constraints for those corresponding regions. This decomposition allows the system to manage complex shapes by controlling individual components independently.
Training-Free Pipeline
SpaceFlow operates without needing additional training data beyond the initial text description and geometric primitives. It achieves local control by enforcing spatial constraints directly within the generative flow process, making it a direct pipeline from input to controllable output.

Terminology

Summary

SpaceFlow presents a training-free pipeline for locally controllable 3D generation from text descriptions and geometric primitives, enabling users to specify whether regions should strictly follow an input shape or allow generative completionEach 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 The gist>

How it works

SpaceFlow is a unified approach for local control of both geometry and appearance during 3D generation In practice, the user specifies the geometry and scene layout as an explicit geometric input, decomposed into an arbitrary set of local parts that can be addressed independently For geometric conditioning, we assign a local control level to each part: higher levels enforce closer adherence to the specified geometry, whereas lower levels leave more freedom to the generative prior The formulation is agnostic to how individual parts are represented, ranging from coarse geometric primitives to detailed meshes We represent parts as superquadric primitives which provide a compact, expressive, and easily editable representation

Local Geometric Control

To implement local geometric control during structure generation, we partition the user-defined primitives S according to their assigned control levels taui First, we convert the complete set S and the high-control subset Shigh into meshes independently We inject the clean high-control latent zhigh1 at each denoising step t Using the low-control mask Mlow, we compute a spatially masked weighted average to update the latent state zt We leverage a RePaint-inspired resampling strategy over K iterations for a smooth transition between high- and low-control regions This iterative resampling is performed across the interval [taulow, tauhigh], corresponding to the Conditioned Flow phase shown in Fig. 2

Local Appearance Routing and Guidance

The appearance stage takes as input the generated sparse structure and the local appearance cues A = ai N i=1, which default to the global text prompt p when unspecified To ensure each local cue affects only its designated primitive S i, we propose local appearance routing We define a routing volume Φ that maps every appearance-latent voxel to its corresponding condition a i, dictating which prompt guides each spatial region along the flow trajectory Constructing Φ requires establishing semantic correspondences between the generated structure and initial primitives We extract PartField [55] descriptors from the generated structure, cluster the active voxels via K-means, and assign each semantic cluster as a unit to its closest input primitive Within every cross-attention layer, each latent voxel attends only to the embedding selected by Φ We inject a soft self-attention bias modulated by Φ This prioritizes intra-part feature propagation without blocking interactions across the rest of the asset

Evaluation and Results

We evaluate the geometric and appearance components separately and demonstrate their combined capabilities on diverse multipart assets Regional geometry metrics show that SpaceFlow preserves strongly constrained regions more faithfully than weak global conditioning, while allowing greater variation in weakly constrained regions than strong global conditioning On fixed geometry, our VLM-asa-judge evaluation of localized text-conditioned appearance shows higher prompt faithfulness and color/material accuracy than the evaluated TRELLIS and GuideFlow3D variants Our quantitative results demonstrate that SpaceFlow reliably follows the high-control input geometry while using the generative prior to match the prompt in low-control regions The regional deltas show that while uniform SpaceControl baselines exhibit virtually no difference in the level of control across regions (ΔCD≈ 0 and ΔD s→10 ≈0), SpaceFlow achieves a substantial positive delta in both physical geometry (ΔCD) and feature space (ΔD s→10)

Limitations

A primary failure mode in our geometric control module occurs when the text prompt contradicts the input geometry Lacking a joint prior for geometric and textual combinations, the model struggles to reconcile both signals Misrouted Appearance Cues occur when a cluster boundary does not coincide with the true material boundary of the generated asset, causing appearance cues to be applied to the wrong voxels Semantic-Geometric Conflicts can lead to unnatural geometric distortions as the model attempts to synthesize the text-described asset around the constrained high-control regions

The gist

SpaceFlow presents a training-free pipeline for locally controllable 3D generation from text descriptions and geometric primitives, enabling users to specify whether regions should strictly follow an input shape or allow generative completion

How it works

SpaceFlow is a training-free framework built upon TRELLIS [3] As shown in Fig.

Improvements for AI systems

  1. Bold Header: Local Geometric Control via Variable Strength Primitives. This capability allows users to specify whether regions should strictly follow the input shape or allow generative completion, enabling a balance between geometric fidelity and generative freedom across an asset, which is superior to uniform global control where it forces a single compromise across the entire asset.

  2. Bold Header: Local Appearance Routing for Part-Specific Conditioning. The system implements a mechanism where appearance cues are routed such that each generated part is conditioned only on its assigned text or image cue, thereby limiting cross-part leakage, which is superior to existing methods where appearance guidance is limited to a single global condition.

  3. Bold Header: Unified Framework for Joint Control. SpaceFlow provides a unified mechanism by combining both approaches, allowing users to specify both geometry and appearance control simultaneously, which enables the user to specify not only what to generate, but precisely where each spatial constraint and appearance cue should apply.

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. The project page is available at SpaceFlow3D.github.io.

Related papers