How does computational photography enhance low‑light shots on modern flagship phones?

08 Ağu 14:52

Modern flagships typically stack three main stages: a burst of short‑exposure frames (often 3‑7 shots), an on‑chip HDR merge, and a final AI‑driven denoiser. The sensor’s readout speed lets the ISP grab a few 1/30 s exposures while the scene is still static enough for alignment. Those raw frames are first combined with a tone‑mapping HDR algorithm that pulls detail from the brightest and darkest captures, effectively widening the dynamic range before any pixel‑level cleaning happens. After the HDR blend, a neural‑network model—trained on millions of low‑light samples—predicts and suppresses residual shot noise, while also sharpening edges that tend to get blurred in the merging step. If you look at the older single‑frame pipeline in, say, the Galaxy S20’s “Night mode” (which relied mainly on a long exposure and basic denoising), the newer multi‑frame approach on the Pixel 8 or Xperia 1 IV delivers noticeably higher signal‑to‑noise and more natural colors, especially when the hand isn’t perfectly steady. The price is extra processing latency (typically 0.8–1.2 s from shutter to preview) and a modest battery hit—roughly a 5‑8 % increase per shot because the ISP has to power the sensor at higher frame rates and run the AI accelerator. Some manufacturers mitigate this by limiting the multi‑frame mode to a few seconds of exposure, or by switching the AI denoiser to a lower‑precision model when the battery is low. So the trade‑off is clear: you get better low‑light quality at the cost of a slight delay and additional power draw, which is why many phones let you toggle “Night mode” on or off depending on your speed vs. quality preference.