ReDT: Resolution-Decoupled Transformer Brings 16x Speedup to HD 3D Reconstruction
zhenjun_zhao · x · 2026-10-23
ReDT is a feed-forward visual geometry transformer that scales multi-view reconstruction to HD inputs, fixing the problem that existing models are trained at fixed SD resolutions and degrade poorly on high-resolution inputs.
The key idea is resolution decoupling: each view is first mapped to a fixed-size set of basis tokens, correspondences are computed in this compact constant-size space, then outputs are projected back to full resolution for dense prediction. This decouples internal attention from input resolution, sharply cutting compute for high-resolution 3D reconstruction.
Across datasets from SD to FHD (2K), ReDT beats SOTA visual geometry transformers in both quality and efficiency — a 16x speedup and 34% lower peak GPU memory on ETH3D (2K).
More from Research
- Berkeley's StyleStream 2.0 cuts streaming voice conversion latency to 520ms, 2x faster — Berkeley · 2026-10-23
- Math PhD student quits academia: AI is making math unrecognizable — pickover · 2026-10-23
- ICLR Submission Found Containing LLM's Own Meta-Comment Admitting Results Weren't Significant — _onionesque · 2026-10-23
- AKV Serving System Lets Agents Manage Their Own KV Cache, Cutting Cost 27.7% — _akhaliq · 2026-10-23
- AuroraTTS-100K: a sub-100K-parameter English TTS running 1120x real-time on GPU — Tall_Abrocoma_3533 · 2026-10-23
- MiTaS multi-resolution tactile imitation learning hits 80% success vs 35.5% vision-only baseline — Jan_R_Peters · 2026-10-23