Nvidia’s 117M-parameter Déjà View rebuilds 3D scenes from photos, research use only

Nvidia has introduced Déjà View, a model that turns photos taken from different angles into 3D geometry and camera positions. By looping one transformer block, the 117M-parameter model “matches or outperforms models up to 10x larger,” the company said.

The post gives no figures beyond that claim. The details below come from Nvidia Research’s project page and the model card on Hugging Face, which spell out how the model works, what it outputs and under what terms it can be used.

One transformer block, applied again and again

According to the project page, Déjà View starts from per-view features produced by a pretrained DINOv2 encoder. It then applies a single transformer block, built from frame-level and global attention, repeatedly to refine its internal state. Two light heads decode the final state into a depth map and a ray map for each view.

The number of repeats is not fixed. Nvidia says it samples that number during training, so one checkpoint can run with any step count inside the trained range. In practice, the repeat count acts as a dial between speed and accuracy at inference time. The authors also report that the looped design beats an otherwise identical variant with separate weights at each step, given matched data and compute.

The model card lists the outputs: per-pixel depth, a six-channel ray map, a point cloud, and camera intrinsics and extrinsics for every view. Input can be RGB images or video in .mov or .mp4 format, decoded into frames. Training used 2 to 18 views per scene at a 504-pixel longest edge, and the card says inference can take more views, limited by memory. Nvidia lists H100 and A100 GPUs as test hardware and says training used 128 H100s.

Who can use it today, and on what terms

The code is on GitHub under nv-tlabs/dvlt, the weights are on Hugging Face as nvidia/dvlt, and the paper is on arXiv (2605.30215). The project page describes the work as a NeurIPS 2026 oral.

The licensing is the main restriction. The weights fall under the NVIDIA License, the Hugging Face tag reads “nvidia-oneway-noncommercial,” and the model card states that the model is for research and development only. The source code is mostly Apache 2.0, with parts derived from Meta’s VGGT under that project’s own license. Developers can experiment with it, but the card does not offer a route to shipping it in a product.

The size comparison needs a closer look

The project page sets Déjà View against Pi3 at 959M parameters, VGGT at 1,257M, Depth Anything 3 Large at 356M and Depth Anything 3 Giant at 1,201M. Against the two largest models, Pi3 and DA3-Giant, Nvidia claims 8 to 10 times fewer parameters and 1.9 to 2.3 times less compute. That compute gap is much smaller than the parameter gap, which fits a design that reuses one block many times: the weights are shared, but the arithmetic still runs once per loop.

On accuracy, the page says the model leads on average inlier ratio and pose AUC@30 degrees across five benchmarks covering indoor, outdoor, object-centric and driving scenes. It supplies no per-benchmark scores and no absolute timings; the charts show only approximate axes. Those results are Nvidia’s own and have not been independently checked, and the arXiv paper is where a reader would look for the full tables.

Other open questions follow from the material. The model card does not say how accuracy changes as the loop count falls toward real-time use, or how the model behaves with far more views than the 18 it saw in training. Neither the post nor the pages say whether a version for commercial use is planned.

A small model in Nvidia’s growing open model catalog

Déjà View sits alongside other Nvidia research and model releases this month. Earlier this week the company published GR00T 2, a robot model with open weights and data, and last month it offered a 4GB Nemotron model for local assistants. Those releases also lean toward compact models. Systems that estimate camera poses and depth feed robotics and mapping work, where a model small enough to run on modest hardware is more useful than a very large one, though Nvidia does not make that link itself.

For now, Déjà View is a research result with usable code and a strong efficiency claim. Whether the looped approach holds its advantage outside the five benchmarks Nvidia chose will become clearer as outside researchers run the released weights.