NeRF learned a scene's geometry from ordinary photographs

The 2020 NeRF paper synthesised new viewpoints from posed photos, at a rendering cost later methods were built to fix.

Historical event

March 19, 2020

First source published

March 19, 2020

Site publication

September 18, 2026

A network that renders viewpoints instead of storing them

On 19 March 2020, researchers posted a paper describing Neural Radiance Fields, later shortened to NeRF, to the arXiv preprint server; it was later accepted to ECCV 2020 as an oral presentation. The abstract states the method achieves "state-of-the-art results for synthesizing novel views of complex scenes" using only "a set of images with known camera poses" as input, ordinary photographs, provided their camera positions have already been calculated by a separate step. No manual 3D modelling, laser scan, or depth sensor is involved. What NeRF actually stores is not a mesh or a point cloud but the weights of a small neural network, described on the authors' project page as representing the scene as "a continuous volumetric function."

The mechanism: a function queried per ray

The network takes a five-dimensional input, a 3D position plus a two-angle viewing direction, and returns a colour and a density at that point. To render a single output pixel, classic volume-rendering mathematics accumulates the network's answers along the camera ray passing through it, and the process repeats for every pixel and every new viewpoint. This is the opposite of a display-ready asset: there is no fixed geometry to hand to a renderer, only a function that must be queried repeatedly, at real computational cost, whenever the viewpoint changes. A NeRF is trained once per scene from a fixed set of photographs and cannot represent a new scene without retraining; it produces a viewing experience, move the camera and the image updates plausibly, rather than an exportable model a display pipeline can simply load.

What the record does and does not establish about display readiness

The original paper demonstrates visual quality and does not claim real-time performance; the abstract's language concerns accuracy relative to prior neural-rendering work, not speed. That the per-ray query cost was a genuine bottleneck for anything resembling a live display is corroborated independently three years later: the paper introducing 3D Gaussian Splatting states plainly that before its own method, "no current method can achieve real-time display rates" at comparable quality and resolution. Read together, the two papers document a specific, dated gap, a scene representation with excellent view synthesis but a rendering cost incompatible with an interactive spatial display, that the field spent years closing. Anyone shown a "NeRF on a headset" demo in 2020 or 2021 was, on this evidence, seeing a research pipeline rather than a shippable display format.

Questions to bring to a demo

• Is the content actually a trained NeRF, or has it already been converted to a faster representation such as a mesh or splats?
• How long did training take for this specific scene, and does that limit how quickly new content can be produced?
• What frame rate is the rendering achieving at the display's native resolution, and on what hardware?

NeRF's contribution was proving that a neural network could stand in for a 3D scene using only photographs. It was never, on the paper's own terms, a display technology; the display-side story is a separate, later chapter the paper itself does not narrate.

Sources & reading trail

States the method, its inputs (posed photographs) and its novel-view-synthesis results; gives the arXiv submission date.

Source published: 19 March 2020 · Retrieved: 16 September 2026

Describes the representation as a continuous volumetric function optimised from images with known camera poses.

Source published: Not established · Retrieved: 16 September 2026

States that prior to this 2023 method, no existing technique reached real-time display rates at comparable quality, corroborating NeRF-era rendering cost.

Source published: 8 August 2023 · Retrieved: 16 September 2026

Primary documents establish the record; the mechanism reading and the demo questions are Presence Atlas editorial analysis. This retrospective draft does not imply the site published on the event date.

Continue reading

• Gaussian splatting rendered radiance fields in real time
• A light field is a claim about how many rays you capture
• A quilt of rendered views feeds a lenticular light field
• Browse the complete the archive

Published September 18, 2026, not on the date of the event described.