SpaceFlow Adds Per-Region 3D Asset Control on TRELLIS

A new paper from ETH Zürich, Stanford and Microsoft, posted to Hugging Face as "SpaceFlow: Locally Controllable 3D Generation," lets a 3D asset generator apply different geometric-adherence strengths and distinct text or image prompts to different parts of the same object, built on top of TRELLIS. The routing uses PartField descriptors and K-means with 30 clusters by default, and blends two denoising paths at τ_high = 10 (strict geometric adherence) and τ_low = 3 (generative freedom).

On an 83-asset evaluation set spanning 7 semantic categories, with assets containing 2-18 parts, SpaceFlow wins 71% of overall-preference votes against SpaceControl at τ=10 and 48% against τ=3, across 337 trials with 29 evaluators. Against TRELLIS it takes 64% of pairwise prompt-faithfulness votes and against GuideFlow3D 61%. On absolute 1-10 ratings, overall scores land at 7.48 for SpaceFlow versus 6.96 for TRELLIS and 7.04 for GuideFlow3D; texture detail is the one axis where TRELLIS edges ahead (7.03 vs 6.48).

The authors frame the motivation plainly: "In practice, asset creation is iterative and part-specific, rather than one-shot: creators establish coarse structure, inspect generated candidates, retain some design decisions, and revise others." They argue that "existing approaches therefore lack a unified mechanism for local control over both geometry and appearance." The listed limitations are concrete: semantic-geometric conflicts when a prompt contradicts the input shape, and misrouted appearance cues when PartField cluster boundaries do not line up with material boundaries. It arrives in a dense week for 3D generation work on our tracker, alongside OuroWorld turning static 3DGS scenes into looping cinemagraphs.