NVIDIA ArtiFixer restores unseen 3D regions with hundreds of frames

Filling the parts a camera never saw

The hard part of sparse 3D reconstruction is not the surface already covered by the camera. It is the under-observed or completely unseen region where otherwise strong 3D methods leave holes, blur, or inconsistent geometry. NVIDIA AI’s June 22 tweet introduced ArtiFixer from NVIDIA Research with a SIGGRAPH 2026 paper, code, and demo.

“fills in the missing geometry that other methods leave blank”

The project page describes ArtiFixer as a two-stage pipeline for improving methods such as 3D Gaussian Splatting when they extrapolate poorly. Existing generative fixes can be expensive because they generate only a limited number of views at once, or they can drift away from the actual scene. ArtiFixer first trains a bidirectional video diffusion model with an opacity mixing strategy so the model stays aligned with observed content while retaining the ability to extrapolate. It then distills that teacher into a causal autoregressive model that can generate hundreds of frames in a single pass.

NVIDIA AI’s account regularly surfaces NVIDIA Research papers, demos, and AI developer work, so this tweet is a technical research pointer rather than a broad product launch. The practical claim is that generated novel views can be used directly or as pseudo-supervision to improve the underlying 3D representation. On the project page, the authors say ArtiFixer3D+ exceeds prior published methods by a wide margin on the challenging three-view split of MipNeRF 360, and beats previous state-of-the-art methods by 1-3 dB PSNR on commonly benchmarked datasets.

The work matters because 3D reconstruction is becoming a base layer for robotics simulation, spatial AI, digital twins, visual effects, and embodied agents. If a model can cheaply fill missing geometry while preserving source fidelity, downstream systems get more complete worlds from fewer captures. The caveat is also important: plausible completion is not the same as measured truth. The next thing to watch is how the released code behaves on messy real scenes, whether temporal consistency holds outside curated demos, and how teams mark generated geometry when accuracy carries operational risk. Source tweet