UofT's Squeeze3D bridges latent spaces to compress 3D data up to 2,187x
rishit_dagli · x · 2026-10-07
Squeeze3D (University of Toronto / Vector Institute, TMLR 2026) compresses 3D data by leveraging implicit priors in pretrained encoders/decoders:
- Method: trains mapping networks converting latent A → unified latent → latent B, without retraining the base models; losses are latent prediction plus a Gram/dimension-wise contrastive term to avoid degenerate dimensions — no reconstruction loss at all
- Results: meshes 2,187x (6.43 MB → 3.01 KB), point clouds 58.5x (117 KB → 2.00 KB), radiance fields 619x (58.07 MB → 0.09 MB)
- Quality: outperforms established codecs like Draco at far lower storage while preserving geometric and scene details
Project page, paper, code and models are all public.
More from Research
- Book anniversary: Data Mining, Practical Machine Learning Tools and Techniques — FrnkNlsn · 2026-10-08
- Bittensor's SN107 Lets Miners Earn by Running AI Agents to Produce Genomic Data — markjeffrey · 2026-10-08
- CoLM 2026 Poster: Vibe-Voting LLMs and Why Users Distrust Benchmarks — boknilev · 2026-10-08
- Baseten's Base Labs has all 3 papers accepted at NeurIPS workshops — baseten · 2026-10-08
- Open-vocabulary text search fused into Google 3D Tiles across 27M voxels of San Francisco — bilawalsidhu · 2026-10-08
- SEAR research project to debut at COLM 2026, paper and code coming soon — WenhuChen · 2026-10-08