VoxLution: sparse voxel cascade trained on real scans cuts 3D-FRONT FPD by 59.6%
angelaqdai · x · 2026-11-03
TUM researchers (Angela Dai's group, with UVA) present VoxLution, a coarse-to-fine sparse autoregressive transformer for high-fidelity 3D indoor scene generation trained on real scans. It refines scenes from 25cm to 1.5625cm voxels, formulates 3D super-resolution as sequence modeling over multi-resolution sparse voxel grids with a training-free resolution-invariant tokenization, and injects adjacent-chunk boundary info into the sequence for coherence. It reports SOTA on 3D-FRONT with 59.6% lower FPD and 51.1% lower FID. Paper and video are out; code coming soon.
More from Multimodal
- Fully local AI music video built with Krea2 visuals and Minimax H3 music — SpiritualLimit996 · 2026-11-03
- Scenario open-sources 3D OS: 80 agent skills that take a sketch to game-ready 3D assets — Scobleizer · 2026-11-03
- Mago Studio open-sources local-first AI creative app running on your own GPU via ComfyUI — TomLikesRobots · 2026-11-03
- "Zero skills" AI artist's Midjourney+Krea short film hits ~200M views in 36 hours — LinusEkenstam · 2026-11-03
- TuneJury: a model trained on human judgments predicts music preferences — mittu1204 · 2026-11-03
- Adobe/MIT: small critics silently break DMD distillation — joint scaling + STAC halve FID — rzhang88 · 2026-11-03