NVIDIA's Déjà View: 117M looped transformer matches models 10x larger at 3D reconstruction
NVIDIAAI · x · 2026-11-12
NVIDIA and partners published Déjà View, a NeurIPS 2026 oral paper on multi-view 3D reconstruction that recycles a single transformer block in a loop instead of stacking unique layers.
- The insight: depth partly buys iteration, paid for inefficiently in unique parameters — so the model applies one block recurrently for K refinement steps, exposing K as an inference-time compute knob with a single checkpoint.
- At only 117M parameters, it matches or beats feed-forward baselines up to 10x larger (π³, Depth Anything 3-G) on five indoor/outdoor/object/driving benchmarks, using 1.9–2.3x less compute.
- Paper, code, and an interactive viewer are publicly available.
More from Multimodal
- 'Harry Potter as a 1973 film' AI video shows retro restyling technique — techhalla · 2026-11-12
- Experts Pan Claude's Beethoven Sonata: 'A Markov Chain on Bach Does Better' — teropa · 2026-11-12
- AI-generated 'Stella Interior' artwork rendered with 2.5 million brushstrokes — VoidStateKate · 2026-11-12
- 15-round editing test: Seedream 5.5 holds scenes far better than GPT Image — kimmonismus · 2026-11-12
- Psychedelic 'Claude Fable 5.2 self portrait' image goes viral on X — repligate · 2026-11-12
- NVIDIA and ElevenLabs expand voice agent partnership with 15MW Blackwell compute and sovereign deployments — nvidia · 2026-11-12