ETH's SpaceFlow: training-free locally controllable 3D generation from text and primitives
ethz · hf · 2026-10-09
ETH Zurich presents SpaceFlow, a training-free pipeline for locally controllable 3D generation from a text description plus geometric primitives. Each primitive proxies an object part with its own control level, letting users choose whether regions strictly follow the input shape or allow generative completion.
- During structure generation, spatial constraints are enforced within the generative flow process
- For appearance, the structure is segmented and matched to primitives; each part is conditioned only on its assigned text or image cue, limiting cross-part leakage
- High-control regions preserve specified geometry while low-control areas vary plausibly; a user study shows competitive overall quality
- On fixed geometry, text-conditioned routing achieves state-of-the-art prompt faithfulness and color/material accuracy
- Project page: SpaceFlow3D.github.io
More from Multimodal
- Reddit: chasing natural AI selfies with Krea 2 Turbo + LoRAs in Forge NEO — Dandypleasure · 2026-10-09
- Reddit: which model for swapping people in video while keeping motion? — SnooStories7879 · 2026-10-09
- Voyager launches as 'Codex for creative work', driving Blender, Resolve and After Effects — rohanpaul_ai · 2026-10-09
- MiniMax H3 ref2va test makes 'Jie Ge' meme characters dance in ComfyUI — luiluiluilammchaou · 2026-10-09
- Qwen Image 2.1 'not great to use' but lovable — user shares working settings — Comfortable-Mind4875 · 2026-10-09
- Filmmaker walks through full OBVPM workflow with MiniMax H3 in ComfyUI — fa6637 · 2026-10-09