IronSight benchmarks 4D reconstruction from two pairs of Meta Ray-Bans

• Bilawal Sidhu's IronSight post shows a 17-second 4D reconstruction prototype made by fusing footage from two pairs of Meta Ray-Bans.
• In a reply about capture, Sidhu said the inputs were “all from the meta glasses,” not a hidden depth rig.
• Sidhu's implementation reply describes the build as “a nice mix of classical and ML,” with a fuller breakdown still pending.
• Sidhu's workflow note says he queued experiments the night before, selected primitives, then built for a few hours the next day.
• Sidhu's follow-up says he also used Fable to finish the roadmap for an older shot-counter project in two nights.

You can watch the IronSight demo, skim Meta's POV capture guide, and compare it with Sidhu's earlier Meta Ray-Ban VPS repo. The research-shaped part is the jump from ordinary POV clips into a navigable time-based scene: the 4D Gaussian Splatting project page frames dynamic scene reconstruction as the problem of fitting a spatio-temporal 4D volume from 2D images.

Bilawal Sidhu, whose public bio lists him as an ex-Google PM for AR/VR and 3D Maps, described IronSight as “a 4D reconstruction” built from two pairs of Meta Ray-Bans. A Google research buddy, according to Sidhu's post, joked that the prototype would have been a SIGGRAPH paper a few years ago.

The clip presents the useful threshold: consumer POV footage is enough to produce a rotating point-cloud reconstruction with time baked into the scene, at least in prototype form. Weekend-build energy is the story here.

The capture stack stayed consumer-grade. When asked whether the footage came from another source, Sidhu's reply said, “Nope, all from the meta glasses.”

Meta's own POV recording guide says Ray-Ban Meta Gen 2 uses an ultra-wide 12 MP camera with a 100 degree field of view and supports up to 3K video. IronSight appears to treat that kind of footage as raw material for reconstruction, not just first-person video.

Sidhu had already been playing in this lane. His earlier see-through-walls repo used smartphones and optional Meta Ray-Ban support with MultiSet VPS, while MultiSet's smart-glasses VPS post describes sub-centimeter 6-DoF localization from existing 3D scans without markers or new infrastructure.

Sidhu has not published the IronSight code or pipeline yet. The public implementation notes are short but concrete:

• Sidhu's workflow note says he queued a batch of experiments the night before.
• The same note says the experiment pass was for testing ideas and selecting the right primitives.
• Sidhu's timing reply puts the next-day build at “just a few hours.”
• His implementation reply calls the stack “a nice mix of classical and ML.”
• That reply says a proper breakdown is coming later.

That “classical plus ML” phrasing matters because 4D reconstruction is still an engineering blend, not a single magic model call. The broader 4D Gaussian Splatting literature describes the target as dynamic view synthesis over space and time, but Sidhu's thread stops before naming the exact primitives IronSight used.

The same toolchain also hit an older creator workflow. Sidhu's follow-up says he used Fable to improve a shot-counter project from last year and finished the full future roadmap in two nights.

The previous version, according to Sidhu's LinkedIn post, used Gemini 2.5 Pro on Meta Ray-Ban footage, combined video and audio to classify hits and misses, and generated an After Effects HUD overlay. Anthropic's Fable 5 announcement describes Fable around long-horizon software engineering, vision, and autonomous coding work, which fits the kind of overnight roadmap compression Sidhu reported.

The aesthetic reference point came from a commenter comparing the look to a point-cloud game. Sidhu's reply answered that the game used the aesthetic beautifully and called IronSight “basically a 4d replay of reality.”

That phrase is the cleanest description of the prototype: not just a scan, not just a video, but a replayable spatial record built from wearable camera footage.