Computational photography is the reason a phone with a sensor smaller than a fingernail can shoot a usable street at night. When you press the shutter, the phone has already been capturing raw frames. It aligns a burst of them (2 to 8 frames in Google's original HDR+, up to 15 in Night Sight), averages them to cut noise, and then tone-maps the result so shadows and highlights both fit on a screen. The "photo" is the output of a pipeline, not one exposure.
- 2 to 8 frames in the original HDR+, up to 15 in Night Sight, are merged per shot.
- √N: merging N frames cuts random noise by the square root of N, so 15 frames give about 2 stops.
- 333 ms per frame handheld and about 1 s braced is Night Sight's exposure cap.
- 0.3 lux is the light level Night Sight was designed to handle.
- Beyond 10× on the Pixel 10 Pro, a diffusion model predicts detail rather than measuring it.
Most explainers stop at "AI makes it better". That hides the useful part.
The core methods are published: Google's HDR+ paper (SIGGRAPH Asia 2016), the Night Sight paper (ACM Transactions on Graphics, 2019) and the handheld super-resolution paper (SIGGRAPH 2019) all give frame counts, exposure caps and timings. Read them and you can predict when your phone will win, when it will smear, and why its photos sometimes look painted.
Why can't a phone sensor just take one good photo?
A phone sensor collects too little light per exposure. Light gathering scales with sensor area, and a typical phone main sensor has a small fraction of the area of a camera sensor. Fewer photons means more shot noise relative to signal, which is a property of light itself, not of bad engineering.
On 27 October 2014, Marc Levoy, then running a Google X imaging team and a professor emeritus at Stanford, published a short post explaining a new HDR+ mode in the Google Camera app for the Nexus 5 and Nexus 6. His own example was a chapel at Princeton. With HDR+ off, the Nexus 6 picked a 1/12-second exposure, and the picture came out both blurred and noisy.
With HDR+ on, the phone instead fired a rapid burst of short exposures over about a third of a second to a second, aligned them starting from the sharpest frame (he called it "lucky imaging") and averaged them. Levoy wrote that he left it switched on, except for fast sport or scenes dark enough to need flash.
Google Research blog: HDR+, low light and high dynamic range photography in the Google Camera app (2014)How small is a phone sensor?
Look at the areas. A 1/1.28-inch type sensor, common in 2025-26 flagships, measures about 10.07 × 7.56 mm, or 76 mm². A full-frame sensor is 864 mm², roughly 11 times larger.
Even the "1-inch" sensors that Xiaomi and others now fit (13.2 × 8.8 mm, 116 mm²) collect about one-seventh of the light of full frame at the same f-number and shutter speed.
| Format | Size (mm) | Area (mm²) | Light vs full frame |
|---|---|---|---|
| 1/1.28-inch type (phone main) | 10.07 × 7.56 | 76 | About 1/11 |
| 1-inch type | 13.2 × 8.8 | 116 | About 1/7 |
| Full frame 35 mm | 36 × 24 | 864 | 1 |
See the data as a table
| Area (mm²) | |
|---|---|
| 1/1.7-inch type | 43.3 |
| 1/1.28-inch type (phone main) | 76.09 |
| 1/1.12-inch type (phone main) | 97.87 |
| 1-inch type | 116 |
| Four Thirds | 225 |
| APS-C (Sony/Nikon) | 368 |
| Full frame 35 mm | 864 |
Source: Wikipedia: Image sensor format (table of sensor dimensions) · Inch-type formats are not standardised, so real sensors vary slightly. APS-C shown at the lower end of its 368-370 mm² range.
Shot noise follows Poisson statistics: if a pixel collects N photons on average, the noise is the square root of N, so the signal-to-noise ratio is also the square root of N. Collect four times the light and SNR doubles. A phone cannot get a bigger sensor into a 9 mm thick body, so it collects more light the other way: over time, across many frames.

How does burst photography (HDR+) work, step by step?
HDR+ captures a burst of deliberately underexposed raw frames at the same exposure, aligns them, merges them into one high-bit-depth image, and then tone-maps that image. Samuel Hasinoff and colleagues at Google described the full system in 2016; it ran in about 4 seconds on the phones of the time and produced a 12-megapixel photo.
- Keep a rolling buffer. The viewfinder already streams raw frames (the paper's viewfinder ran at 15-30 frames per second). Google's later zero-shutter-lag mode sends the most recent 9 or 15 frames to the merge, so the burst partly happened before your finger landed.
- Underexpose on purpose. Every frame uses the same short exposure, set low enough that highlights do not clip. The paper caps this deliberate underexposure (its "dynamic range compression") at a factor of 8, three stops, because more makes the final image look cartoonish.
- Start from raw, not JPEG. Bayer raw frames from the sensor are typically 10 bits per pixel against the 8 bits of the processed YUV stream. The authors note the true gain is under 2 bits, because raw is linear, but it also lets them skip the hardware chip's own tone mapping and denoising.
- Align tiles. Each frame is cut into tiles and matched to a reference frame with a coarse-to-fine search, so hand shake is undone tile by tile.
- Merge robustly. Tiles are combined in the frequency domain with a Wiener-style filter that averages where frames agree and falls back to the reference where something moved. That is why a walking person usually comes out sharp but slightly noisier than the wall behind.
- Finish the image. Demosaic, white balance, local tone map, dehaze, sharpen and convert to the final JPEG or HEIC.
Burst length is a trade. In bright light the paper found 1-2 frames usually enough; in dim or very high-contrast scenes it wants more. In practice it limited bursts to 2-8 frames, chosen by a noise model. Exposure settings followed a fixed schedule: hold gain at minimum and let exposure rise to 8 ms, then hold 8 ms and raise gain to 4×, then raise both up to 100 ms and 96× gain.
The Google team describes the HDR+ pipeline as shipped on mass-market phones. It captures 2-8 constant-exposure raw frames, aligns them with a tile-based method, merges them with a hybrid 2D/3D Wiener filter and tone-maps the result, all in about 4 seconds on device for a 12 MP output. Auto-exposure is tuned against a hand-labelled database of about 5,000 scenes, and quality was checked on 26,071 real scenes. The limit: it is a systems paper about a commercial product, so it gives design choices and timings rather than a controlled comparison with other phones.
Read the paper (free)Why does merging frames reduce noise?
Averaging N frames of the same scene cuts random noise by the square root of N, because the signal adds up linearly while uncorrelated noise adds up as a square root. Eight frames give about 2.8 times the SNR of one frame, which is 1.5 stops; fifteen frames give 3.9 times, or almost 2 stops.
Calculated in this post from the noise model in Hasinoff et al. (2016), where averaging N frames lowers the noise variance to σ²/N, so SNR rises by √N: √8 = 2.83 and √15 = 3.87, which are log₂(2.83) = 1.5 and log₂(3.87) = 1.95 stops.
That gain is the whole trick. A phone that merges 15 frames behaves, for noise, roughly like a sensor with 15 times the area capturing one frame of the same length, provided nothing moves.
Motion is what breaks the analogy: each frame that cannot be aligned is rejected locally, so moving parts of a scene get fewer frames and stay noisier.
What a burst buys you
What does Night Sight do differently from HDR+?
Night Sight keeps the burst-and-merge idea but stretches each frame much longer and decides the length by measuring motion first. HDR+ caps viewfinder frames at about 66 ms; Night Sight lets each frame run to about 333 ms handheld and up to about 1 second when the phone detects it is braced or on a tripod.
Google's Night Sight write-up and the Liba et al. paper give the operating limits:
| Mode | Frames | Max exposure per frame | Example total | Usable light level |
|---|---|---|---|---|
| Single-shot phone capture | 1 | Short (scene-dependent) | One frame | Struggles below about 30 lux |
| HDR+ (viewfinder burst) | 9 or 15 recent frames sent to merge | About 66 ms | Under 1 s | Works down to about 3 lux |
| Night Sight, handheld | Up to 15 | About 333 ms | 15 × 333 ms = 5 s | About 3 lux down to 0.3 lux |
| Night Sight, braced or tripod | Up to 6 | About 1 s | 6 × 1 s = 6 s | Autofocus fails below about 0.3 lux |
Motion metering
The clever part is motion metering. Before capture, the phone estimates how much the scene and your hand are moving, then picks the frame count and per-frame time that balance blur against noise. In Google's example, a scene metered without motion detection would have used 73 ms frames; with it, 48 ms. Shorter frames mean each one is sharper, and the merge recovers the noise.

Thirteen Google researchers describe Night Sight: motion metering to set frame count and exposure, a merge designed for very noisy frames, a learned auto white balance and a tone map that deliberately keeps night looking dark. The system produces photos from a single shutter press in a few seconds, handheld, in light as low as 0.3 lux, running entirely on the phone. The limit: the evaluation leans on visual comparisons and Google's own devices, and the motion metering cannot save subjects that move fast during a multi-second capture.
Read the paper (free)Why do phone night photos look brighter than reality, and sometimes painted?
Because the tone map decides how bright night should look, and the denoiser decides which texture is real. Both are aesthetic choices built into the pipeline, so a phone photo reflects the manufacturer's taste as much as the scene.
The Night Sight paper is unusually honest about this. Its tone mapping borrows from illusionistic painting: it adds contrast, lets shadows fall to black and keeps the frame edges dark, so the result reads as night even though the merged data could be brightened to look like day.
The merged data could be brightened to look like day; Google chose to keep night looking like night.This post, summarising Liba et al., ACM Transactions on Graphics, 2019
Other makers push brighter. When people call phone photos "over-processed", they are usually reacting to three things:
- Aggressive local tone mapping. HDR stacking captures a wide range, then compresses it so the sky and the shadowed face both sit mid-scale. Push that too far and you get flat, halo-edged pictures. The top answer on a long-running Stack Exchange thread makes the point well: a high-range image squeezed onto a normal screen looks flat, and local tone mapping is how you win the contrast back, region by region.
- Denoising that eats texture. Skin, foliage and distant brick have fine random detail that looks statistically like noise. A strong denoiser smooths it, then sharpening adds crisp edges back, giving a watercolour look.
- Semantic adjustments. Modern phones segment the frame into sky, skin, food and text, and process each region differently. That can make a sky bluer than the eye saw it.
White balance matters too. Under sodium or mixed LED light, a learned white balance chooses what "neutral" should mean, which is why the same street can look orange on one phone and cool on another. I cover how the eye and camera disagree about this in white balance and colour temperature.
How does Super Res Zoom get more detail than the sensor has?
It uses your hand tremor. Each frame of a handheld burst lands on the sensor shifted by a fraction of a pixel, so together the frames sample the scene at positions a single frame never sees. Combining them gives real extra resolution and replaces the usual demosaicing step.
A normal sensor sits under a Bayer colour filter: each pixel sees only red, green or blue, and two-thirds of the colour at every point is guessed by interpolation (demosaicing).
Wronski and colleagues showed that a burst of raw frames, aligned with sub-pixel accuracy, can fill in true red, green and blue at every point. Their method processed each 12-megapixel raw frame in about 100 ms on a phone, and became both Super Res Zoom and the default merge in Night Sight.

The authors replace demosaicing with a multi-frame merge that builds a full RGB image straight from a burst of raw colour-filter-array frames, using natural hand tremor as the source of sub-pixel offsets. The merge uses kernel regression with a robustness model that handles local motion, occlusion and scene change. It runs at about 100 ms per 12 MP raw frame on a mass-market phone. The limit: on a tripod there is no tremor, so the method loses its free sub-pixel offsets, and fast motion still reduces the gain.
Read the paper (free)Where does generative AI start, and is it still a photo?
Generative zoom starts where the optics end. Google's Pro Res Zoom on the Pixel 10 Pro, explained by Google in October 2025, uses multi-frame Super Res Zoom up to 10× and then a single-step diffusion model on the Tensor G5 chip for 10× to 100×. That model is asked to generate detail the sensor did not record, guided by the shapes and colours it did.
This is a different kind of processing. Burst merging only rearranges photons that arrived; nothing in an HDR+ image is invented, apart from aesthetic decisions about tone.
A diffusion model predicts plausible texture, and plausible is not the same as true: reviewers have found misshapen animals and missing parts at extreme zoom.
Do bigger sensors and 200 MP still matter if software does the work?
Yes, because every step above works better with more light per frame. A bigger sensor gives each frame a higher starting SNR, so the phone needs fewer frames, rejects less motion and denoises less. That is why 2026 flagships are growing sensors, not just models.
At MWC Barcelona in March 2026, RedShark News listed two flagships built around bigger sensors:
- Xiaomi 17 Ultra: a 1-inch, 50 MP main sensor and a 200 MP 1/1.4-inch telephoto;
- Vivo X300 Ultra: a 200 MP Sony LYT-901 main sensor (1/1.12-inch) plus a 200 MP periscope.
Very high pixel counts mostly serve cropping and binning: a 200 MP sensor typically combines groups of pixels into one in low light. Whether those megapixels show up as detail is a separate question, which I take apart in do megapixels matter. For how current phones score in a standard test, see phone cameras ranked by DXOMARK.
How do you work with the pipeline instead of against it?
Hold still for a second after the shutter sound, keep moving subjects out of night shots, and shoot raw when you want to make the tone decisions yourself. The pipeline is built around a few assumptions, and meeting them gives you cleaner files.
- Brace the phone. Night Sight raises per-frame exposure from about 333 ms to about 1 s when it detects stability, and it can use fewer, longer frames. A wall or railing helps more than any setting.
- Let the burst finish. Some capture continues after the press. Moving early blurs the reference frame.
- Avoid fast subjects in very dark scenes. A rider passing a streetlight becomes a ghost or a noisy patch, because only one or two frames align there.
- Use raw for high-contrast scenes. Phone raw formats (DNG) usually carry the merged, high-bit-depth result before the maker's tone map, so you get the stacking benefit without the "painted" look. The trade-offs are the same as on a camera, covered in shooting raw for dynamic range.
- Turn off extreme zoom for evidence. If the picture needs to be true, such as a signboard or a number plate for an insurance claim, stay within optical and Super Res range.
For the practical side of composing and shooting with a phone, see the phone photography guide. I use a Samsung Galaxy S25+ for extra coverage next to dedicated cameras, and the rules above are the same on any modern phone. The rest of my kit is on the gear page.

What's new
- Google explains Pro Res Zoom: a diffusion model on Tensor G5 takes over beyond 10× up to 100× Google Blog, Oct 2025
- MWC 2026 brings 1-inch main sensors and dual 200 MP cameras to Xiaomi and Vivo flagships RedShark News, Mar 2026
Questions people ask
Is tone mapping the same as HDR?
No. HDR is capturing a wider range of brightness than one exposure can hold; on a phone that is done by merging a burst. Tone mapping is squeezing that range into what a screen or print can show. The "HDR look" people complain about is a tone-mapping choice, which is why a Stack Exchange thread on the topic separates the two.
Why do my phone's night photos look brighter than what I saw?
Because merging up to 15 frames of about a third of a second each collects several seconds of light, then the tone map chooses a brightness. Google deliberately darkened Night Sight's output so night still looks like night; other makers choose brighter renderings.
Does computational photography make megapixels irrelevant?
Not irrelevant, but secondary. Merging fixes noise and some resolution loss, while pixel count mainly helps cropping and binning. Sensor area and lens quality still set how much real light and detail each frame starts with.
Can a phone light meter app replace a handheld meter?
For reflected-light metering in ordinary light, often yes. On Stack Exchange's photography site, one user checked an iPhone meter app against a Canon 5D Mark II and film results and found it accurate, and the developer of another app claimed agreement within a tenth of a stop. Both note the limits: no incident-light reading without a diffuser dome or add-on, and poor performance in very dim light, where phone sensors are already near their single-frame limit.
Is a Pro Res Zoom photo a real photograph?
Up to the multi-frame range, yes: it is merged sensor data. Beyond 10× on the Pixel 10 Pro, a diffusion model predicts missing detail, so treat the fine texture as a reconstruction rather than evidence.
Sources
- Hasinoff et al.: Burst photography for high dynamic range and low-light imaging on mobile cameras, SIGGRAPH Asia (2016)
- Liba et al.: Handheld Mobile Photography in Very Low Light, ACM TOG 38(6) (2019)
- Wronski et al.: Handheld Multi-Frame Super-Resolution, SIGGRAPH (2019)
- Marc Levoy, Google Research blog: HDR+, low light and high dynamic range photography in the Google Camera app (2014; source for the story)
- Google Research: Night Sight, seeing in the dark on Pixel phones (2018)
- Google Research: HDR+ publication page (2016)
- Google: See Pixel 10 Pro's new zoom tech in action (2025, updated 2026)
- Wikipedia: Image sensor format, sensor dimensions table (2026)
- RedShark News: All the major smartphone camera news from MWC Barcelona 2026 (2026)
- Photography Stack Exchange: What is tone mapping? How does it relate to HDR?
- Photography Stack Exchange: Do megapixels matter with modern sensor technology?
- Photography Stack Exchange: How well do smartphone light meter apps work?

