Impossible Camera logo

SEE IN THE DARK

IMPOSSIBLE CAMERA

An iOS camera that builds a photograph from the photons most cameras throw away.

The idea

A modern phone sensor can hear the room, but only if you give it time. Impossible Camera takes that time and makes it work harder. We pair extended exposure with a denoising pipeline trained on millions of paired low-light and ground-truth frames, so the sensor isn't guessing what you saw, it is reconstructing it.

The work happens on the device. No round trip to a server, no pictures leaving your phone. The model is small enough to run alongside the camera UI, fast enough that the preview is the photograph, and quiet enough that you forget it is there.

It was built for the moments traditional cameras give up on. Streetlight at four in the morning, a dim restaurant, a bedroom with a curtain pulled across. Astrophotography mode pushes further, stacking minutes of exposure into a single image. The point is not novelty. The point is that the picture you remember and the picture you take become the same picture.

Capabilities

WHAT IT DOES

Photon reconstruction

Recovers detail from scenes lit by a single candle (around 0.05 lux), with noise floors below the sensor's own shot noise.

Fully on-device

Inference runs locally on the Apple Neural Engine. Photos never leave your phone, even briefly.

Manual and auto

Full ISO, shutter and white-balance control, or one-tap capture with sensible defaults for everyday use.

RAW pipeline

Save the reconstructed image as a 14-bit DNG for editing in Lightroom, Capture One or Halide.

Astrophotography mode

Multi-frame stacking for skies, with built-in star alignment and trail compensation.

Live preview

The viewfinder shows the reconstructed image in real time, not the raw sensor feed.

Under the hood

HOW IMPOSSIBLE CAMERA WORKS

Three stages, from photons to picture.

  1. 01

    Capture

    A long-exposure burst with adaptive ISO. The sensor records the noisiest version of the scene that still preserves edge information.

  2. 02

    Reconstruct

    A U-Net class denoiser, trained on paired low-light and ground-truth frames, separates signal from sensor noise, photon shot noise and read noise.

  3. 03

    Render

    Tone mapping returns the image to natural latitude. Color is preserved through a chromatic prior trained on calibrated film and digital references.

Beta

JOIN THE BETA

Camera is in private beta. Add your name to the list, we are letting people in steadily.