Chapter 23 — Key Takeaways

The one-page reference for computational photography. Screenshot this before a shoot. The big idea in a sentence: a modern phone computes a photograph from many frames rather than recording one — so the skill is knowing when to let the computation work and when to override it.


The core terms

Term Plain-language definition
Computational photography Photography where the final image is produced by software combining and processing data from one or more captures, not by a single recorded exposure (introduced Ch.22).
Image stacking Capturing multiple frames of the same scene and merging them pixel-by-pixel for less noise, more dynamic range, or more sharpness. The foundational move.
HDR (high dynamic range) Capturing several exposures and merging them to hold a scene's full brightness range (e.g. shadowed face and bright window).
Tone mapping Compressing a wide captured brightness range down to fit a display/print — deciding which tones to lift, hold, and crush. Gentle = natural; heavy = the "HDR look."
Depth map A computed image recording distance (not color) at each pixel — how far each part of the scene is from the camera.
Synthetic bokeh A digitally simulated background blur applied by the depth map, imitating a fast lens (portrait mode).
Semantic segmentation The phone identifying what regions are (sky, face, foliage, food) and processing each with a separate tuned recipe.

Why many frames beat one (the engine)

SIGNAL (real brightness) is the SAME every frame  → reinforces when averaged → survives
NOISE (random speckle)   is DIFFERENT every frame → averages toward zero     → shrinks
  More aligned frames → cleaner, higher-range, sharper image.
  REQUIRES: the scene holds still long enough to be "borrowed from" across time.

The four computational modes at a glance

Mode What it does Home territory (trust it) Fatal weakness (override it)
Multi-frame / stack Averages a burst to cut noise Any dim scene; runs invisibly even in "normal" Motion between frames
HDR Merges exposures for range, then tone-maps High-contrast: backlit, bright windows, bright skies Mood/low-key shots (opens shadows you want dark)
Night mode Aligns + rejects + averages a burst in the dark Static dark scenes: empty streets, lit buildings, still life Moving subjects (smear/ghost)
Portrait mode Builds a depth map, fakes shallow DoF Subject far from a simple background, moderate blur Hard edges: hair, glasses, reaching hands, clutter

How to shoot each mode well — checklists

Any multi-frame capture - [ ] Brace hard (both elbows, lean on something) — the frames must align. - [ ] Tap to set exposure on what matters before you press — you're aiming the computation. - [ ] Hold still until the thumbnail settles — moving away early leaves it half-merged.

HDR - [ ] Use it for high-contrast scenes; skip it for mood/low-key (protect the shadows). - [ ] Tap to expose for the highlights; let the merge lift the shadows (blown highlights can't recover). - [ ] Prefer "natural" over "rich/vivid"; watch for edge halos (sign of overcooked tone mapping).

Night mode - [ ] Brace harder and longer than feels necessary; a ledge or mini-tripod extends the capture. - [ ] Drag exposure down so streetlights/neon/the moon keep detail and color. - [ ] Tell any subject to freeze for the whole capture — or add light for a shorter capture.

Portrait mode - [ ] Step the subject forward, away from the background — real distance = easy depth map. - [ ] Use moderate blur, not maximum (maximal exaggerates every seam). - [ ] Mind the four traps: hair, glasses/transparent things, reaching hands/objects, cluttered close bg. - [ ] Inspect the edge at 100%; if it's mangled, fix the scene, don't max the slider.


The "phone look" pipeline (runs on nearly every frame)

Stage What it does The lie to watch for
Demosaic Reconstructs color from the sensor mosaic
Stack / HDR / Night Merges frames (if engaged) Ghosting if something moved
White balance + color Neutralizes, then boosts "memory" colors Bluer skies / warmer skin than reality
Local tone + contrast Opens shadows, adds punch Flat grey "HDR look," edge halos
Sharpening Exaggerates edges for crispness Crunchy, brittle, haloed edges
Noise reduction Smooths speckle Waxy skin, mushy foliage — the "watercolor" smear
Segmentation tweaks Sky/face/foliage/food each tuned separately Confidently "fixes" a sky/face you wanted as-is

Bottom line: the processed JPEG is an interpretation, not a record.


The override decision — when computation lies

Scene / goal Computation's instinct Do this instead
Moody / low-key / silhouette HDR opens every shadow → flat HDR off; one dark exposure; protect the dark (Ch.8)
Moving subject in the dark Night mode smears/ghosts Subject freezes, or add light, or single brighter frame
Hard portrait (hair/glasses/clutter) Portrait mode mis-segments Real distance; moderate blur; or skip the fake blur
Accurate color needed Memory-color boost drifts Shoot RAW; set white balance yourself
Fine texture matters Noise reduction smears it Shoot RAW; control NR in edit; more light at capture
Keeper / print / pro job Pipeline bakes in irreversible choices Shoot RAW; you make the decisions (Part VI)

The master override: RAW vs. JPEG

Computed JPEG/HEIC RAW
Out of camera Gorgeous; HDR/night magic baked in Flat, unimpressive, no auto-magic
Control Choices locked in, partly irreversible Full control over color, tone, texture, NR
Latitude Limited (8-bit, processed) Maximal highlight/shadow recovery
Truth An opinion about the scene The real sensor data
Best for Speed, sharing, good-light snaps Keepers, accuracy, hard light, prints, editing
Pro move Shoot RAW + JPEG — lose nothing but storage

Top common mistakes → fixes

Mistake Fix
Jerking the phone away the instant you hear the shutter Hold until the thumbnail settles — the capture is still running
Auto-HDR flattening a mood shot Turn HDR off; expose for one dark frame; keep the shadows dark
Night mode on a moving subject, then blaming the dark Subject freezes / add light / single exposure — match tool to motion
Maxing portrait-mode blur on a busy background Moderate blur + step the subject forward for real distance
Trusting the JPEG as "what the camera saw" Shoot RAW when accuracy or fine control matters
Brighter light to fix "plastic" skin Often makes it worse; shoot RAW and dial noise reduction down

Quick principles (the prose 20%)

  • Stacking turns time into quality. A tiny sensor borrows in time the light it lacks in area — but only when the scene holds still long enough to borrow from.
  • HDR's merge recovers range; its tone-map decides whether the photo looks real or cooked. Two phones shoot the same sunset differently because their tone-mapping taste differs, not their physics.
  • Portrait mode substitutes math for an optics limit a small lens genuinely can't overcome — so it can never be perfect; it's estimating 3-D from a flat image.
  • Computation rewards a scene that's already easy to read. Give it distance, light, and stillness, and it looks like magic; feed it motion, clutter, and hard edges, and the seams show.
  • The machine supplies the measurement; only you supply the meaning. It can't see what the photograph is about.

Portfolio Checkpoint (this chapter's increment)

Assignment: Reshoot a tricky scene (backlit/HDR, dark/Night, or needs-separation/Portrait) with a computational mode — after shooting the honest single-exposure baseline first. Keep the best computational frame.

Curation note (the real point): Beside the image, write two lines — 1. What the computation added (range you couldn't hold; noise it cleaned; a blur a phone lens can't make). 2. *What it faked* or got wrong (a hair halo; a too-flat tone-mapped sky; a smoothed face; a snapped cutout; a pushed color).

This frame proves you can make a computational camera serve your vision — and see through its tricks.


Spaced review hooks (from earlier chapters)

  • Ch.22: Why a small sensor struggles in low light; bracing, adding light, exposure lock.
  • Ch.2: RAW vs. JPEG — what each holds, and why it's the key to overriding the pipeline.
  • Ch.3: Dynamic range and the histogram — predicting what a single exposure will clip.
  • Ch.5 / Ch.11 / Ch.4: Light direction, fill, and focus — the fundamentals computation sits on top of.