---
type: "post"
id: "nvidia-released-deja-view-a-117m-parameter-multi-8bc008"
slug: "nvidia-released-deja-view-a-117m-parameter-multi-8bc008"
title: "NVIDIA released Déjà View, a 117M-parameter multi-view 3D reconstruction transformer, NeurIPS 2026 oral."
description: "NVIDIA Research describes Déjà View, a 117-million-parameter transformer that reconstructs camera poses and dense scene geometry from multiple views by repeatedly applying one shared block."
retrieval_nugget: "NVIDIA Research describes Déjà View, a 117-million-parameter transformer that reconstructs camera poses and dense scene geometry from multiple views by repeatedly applying one shared block. Its reported results suggest that explicit iterative refinement can compete with much larger models while letting users trade inference compute for refinement steps—a useful research direction for teams handling multi-view imagery."
published_at: "2026-11-12"
updated_at: "2026-11-12"
record_date: "2026-11-12"
date_kind: "published_at"
topics: ["multimodal","models"]
entities: ["NVIDIA","Déjà View","DINOv2","NeurIPS 2026","π³","Depth Anything 3","VGGT","Pi3"]
related_patterns: []
novelty_label: "new_fact"
what_changed: "NVIDIA introduced Déjà View, a 117M-parameter model for multi-view 3D reconstruction that lets users adjust inference compute and was accepted as a NeurIPS 2026 oral."
prior_refs: []
source_url: "https://research.nvidia.com/labs/dvl/projects/dvlt"
source_urls: ["https://research.nvidia.com/labs/dvl/projects/dvlt"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:nvidia-released-deja-view-a-117m-parameter-multi-8bc008"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/nvidia-released-deja-view-a-117m-parameter-multi-8bc008/","markdown":"https://newruntime.com/posts/nvidia-released-deja-view-a-117m-parameter-multi-8bc008.md","json":"https://newruntime.com/posts/nvidia-released-deja-view-a-117m-parameter-multi-8bc008.json"}
markdown_url: "https://newruntime.com/posts/nvidia-released-deja-view-a-117m-parameter-multi-8bc008.md"
---

# NVIDIA released Déjà View, a 117M-parameter multi-view 3D reconstruction transformer, NeurIPS 2026 oral.

## Retrieval answer

NVIDIA Research describes Déjà View, a 117-million-parameter transformer that reconstructs camera poses and dense scene geometry from multiple views by repeatedly applying one shared block. Its reported results suggest that explicit iterative refinement can compete with much larger models while letting users trade inference compute for refinement steps—a useful research direction for teams handling multi-view imagery.

## Summary
NVIDIA Research describes Déjà View, a 117-million-parameter transformer that reconstructs camera poses and dense scene geometry from multiple views by repeatedly applying one shared block. Its reported results suggest that explicit iterative refinement can compete with much larger models while letting users trade inference compute for refinement steps—a useful research direction for teams handling multi-view imagery.

## What happened
NVIDIA Research’s project page presents Déjà View as a method for multi-view 3D reconstruction. Given multiple input views, it estimates camera poses and consistent depth, from which scene geometry can be reconstructed. The page identifies the work as “Déjà View: Looping Transformers for Multi-View 3D Reconstruction,” lists Alessandro Burzio and Tobias Fischer as equal contributors, and names researchers affiliated with NVIDIA, the University of Modena and Reggio Emilia, the University of Toronto and Vector Institute, and ETH Zürich. The page lists NeurIPS 2026 (Oral) and links to a paper, code, and an interactive viewer.

The architectural choice is to make repeated refinement explicit. The system initializes per-view features using a pretrained DINOv2 encoder, then applies one transformer block recurrently. That block has frame-attention and global-attention sub-blocks; each application is conditioned on a continuous time interval. During training, the number of refinement steps, K, is sampled from a range. NVIDIA says this lets one trained checkpoint cover any step count in that range at inference, making K a compute control rather than a choice that requires a separately trained model. Two lightweight heads decode the final state into per-view depth and ray maps.

The project page reports 117 million parameters. Across five reconstruction benchmarks spanning indoor, outdoor, object-centric, and driving scenes, NVIDIA says Déjà View matches or surpasses substantially larger feed-forward baselines. It reports 8–10 times fewer parameters and 1.9–2.3 times less compute than π³ and Depth Anything 3 — G in the stated comparison. The page also says Déjà View leads the average inlier ratio and pose AUC@30° across all five benchmarks, while having the smallest parameter count among the baselines shown. Those comparisons include systems such as Pi3, VGGT, and Depth Anything 3 variants, whose parameter counts are displayed on the page.

NVIDIA further reports that the shared-block version outperforms an otherwise identical variant with independent per-step parameters under matched training data and compute. The interactive viewer exposes intermediate refinement states and lets visitors inspect reconstructions and compare them with feed-forward baselines. The page does not state a product price or describe a hosted service; it links to code and a paper for readers who want to examine the method.

## What the archive adds
New Runtime’s earlier coverage of NVIDIA’s Nemotron 3.5 Super VL discussed an open multimodal model for specialized agent workflows ([earlier coverage, 21 October 2026](https://newruntime.com/posts/nvidia-released-nemotron-3-5-super-vl-an-3a7957/)). That is a separate model and task, not prior coverage of 3D reconstruction. Déjà View therefore opens a new thread in New Runtime’s coverage rather than adding a fact to that story.

## Trend signal
This does not yet fit a pattern New Runtime tracks. Déjà View is a research result about iterative inference in a vision model; it does not establish a broader pattern in agent systems, software interfaces, or generative-media applications.

## Why now
The project page lists the work as a NeurIPS 2026 oral. Its publication framing makes the result timely for readers following research in multi-view reconstruction, while the interactive viewer offers a way to inspect the authors’ examples alongside the reported benchmark results.

## Why it matters
The relevant design question is not simply whether a larger model can reconstruct a scene more accurately. Déjà View tests whether some of the gains associated with depth can instead come from repeatedly applying a shared block, and whether users can control the amount of that iteration after training. If the reported comparisons hold up, this could give teams a more direct way to budget inference for reconstruction workloads: spend more compute on difficult scenes or reduce it where speed matters more.

The result is not yet a deployment recommendation. Benchmarks do not tell a team how the method behaves on its own cameras, image quality, motion patterns, or failure cases. But the architecture and viewer make a concrete evaluation question possible: does increasing K improve the geometry that a downstream workflow actually needs, enough to justify the added computation? That is more actionable than treating parameter count alone as a proxy for cost or usefulness.

## Limits and uncertainty
The performance and compute comparisons on the project page are claims by the research authors; they have not been independently checked here. The page reports aggregate results across five benchmarks but does not, in the cited summary, quantify how performance varies by scene type or refinement-step setting. It also does not establish whether the method is robust to real-world capture problems such as blur, occlusion, or inconsistent lighting. The page links to code and a paper, but no commercial availability, support commitment, or pricing is stated there.

## Skill to build now
If your team already reconstructs scenes from multiple images, use the project’s viewer and linked implementation as a shortlist for a controlled comparison. Pick representative scenes, record reconstruction quality and runtime at multiple refinement-step counts, and compare those results with your current baseline. Keep the same images and evaluation criteria across systems; the practical question is whether extra iterations improve a downstream task, not only whether a visualization looks sharper.

## Reader takeaway
Déjà View is a research demonstration of a useful architectural trade-off: one shared transformer block can be looped to refine a 3D reconstruction, with inference compute exposed through the number of steps. NVIDIA reports strong results from a relatively small model, but teams should treat the benchmark claims as a reason to test the approach on their own scenes—not as proof that it is already the best production choice.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Multimodal](https://newruntime.com/topics/multimodal.md): Explore the multimodal topic hub.
- topic hub: [Models](https://newruntime.com/topics/models.md): Explore the models topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)