SpaceFlow: Locally Controllable 3D Generation
AuthorsNeil De La Fuente, Joan Lafuente, Mukhammadali Sayfiddinov, Felicia Scharitzer, Marc Pollefeys, Ata Celen, Sayan Deb Sarkar, Elisabetta Fedele
Affiliations[
SpaceFlow lets creators control how closely individual 3D regions follow supplied shapes and independently guide each part’s appearance.
Key results
Assets used to evaluate local geometry and appearance control.
SpaceFlow win rate against SpaceControl with global control strength τ=3 on prompt-and-geometric fidelity.
SpaceFlow win rate against SpaceControl with global control strength τ=10.
SpaceFlow mean score on the 1–10 text-conditioned appearance scale.
SpaceFlow mean score on the 1–10 text-conditioned appearance scale.
SpaceFlow mean score on the 1–10 text-conditioned appearance scale.
What the paper found
SpaceFlow is a training-free 3D generation framework built on Microsoft’s TRELLIS that lets users control geometry and appearance independently for different object parts. Editable superquadric primitives receive local geometry-control strengths: masked latent blending and RePaint-inspired resampling preserve high-control shapes while allowing low-control regions to be completed by the generative model. For appearance, PartField clusters match generated regions to the input primitives, then routed cross-attention sends each part only its assigned text or image cue; a soft self-attention bias and contrastive latent guidance help maintain coherent results. Text cues use OpenAI’s CLIP, while image cues use DINOv2. On an 83-asset benchmark, SpaceFlow beat global low-control guidance on prompt-and-geometric fidelity in 78.3% of comparisons, and a user study preferred it over global high-control guidance in 71% of trials. On a 1–10 appearance scale, it scored 7.07 for prompt faithfulness, 6.43 for color and material accuracy, and 6.78 overall, outperforming the evaluated appearance baselines on overall quality while trading off some texture detail.
Original 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.
Read the original paperMore in Computer Vision
Browse all 67 papers →CLoSeR: Closing the Loop for Long-Context Streaming Reconstruction
Moyang Li, Zihan Zhu, Wei Zhang, Marc Pollefeys, Daniel Barath
CLoSeR helps 3D reconstruction models stay accurate over kilometer-long streams by detecting revisited locations and correcting accumulated pose drift.
CompactSplat: Spatially Adaptive Gaussian Distribution for Feedforward Scene Reconstruction
Bing He, Jingnan Gao, Yunuo Chen, Zhengxue Cheng, Rong Xie, Li Song, Wenjun Zhang
CompactSplat reconstructs 3D scenes by placing Gaussian primitives densely where detail matters and sparsely across flat regions.
LEGO: A Lifting-Free Approach for Exocentric-to-Egocentric Video Generation
Suhwan Cho, Yonwoo Choi, Soongjin Kim, Jicheol Park, Taegyu Lim
LEGO replaces brittle depth-based camera lifting with a learned transformer view synthesizer that gives diffusion models better structure for generating first-person video.