Case Study 33.2 — Shoot-Along: The Honest AI Pipeline (and the Comparison Pair)
"Above all, life. A photographer must never lose this curiosity for the everyday." — a sentiment widely attributed to documentary photographers
Case Study 33.1 was analysis: a single witnessed image, taken apart. This one is production — a complete, from-scratch shoot you can run today at your own kitchen window, followed by an edit that uses AI tools honestly, staying on the editing side of the line, and ending with the disclosed "editing vs. generating" comparison pair (Figure 33.4) that you can put straight into your Portfolio. The goal is not a pretty picture. It is to feel, with your own hands, exactly where the editing/generating line is, what honest AI assistance looks like end to end, and how to disclose it.
The brief and the constraints
The brief: produce one finished still-life photograph of an ordinary object at the kitchen window (Chapter 13's soft, controllable light), deliberately shot in conditions noisy enough that AI denoise has real work to do — then take it through an honest pipeline, and build a teaching comparison showing the same capture finished by editing versus generating.
The constraints, chosen on purpose: - One subject, one window. A single humble object — we'll use a chipped enamel mug, but a piece of fruit, a shell, a pair of glasses all work. Anyone can do this on foot, indoors, today. - Shoot it dim and late. We will photograph near dusk with the window light fading, deliberately driving the ISO high so the file is genuinely noisy. That gives the AI denoiser a real job and lets you see exactly where restoration ends and invention begins. - One honest rule, declared up front: on the editing side, we may denoise, mask for local tonal work, grade, crop, and modestly upscale — all of which refine what the light recorded. We will not add or remove any content in the portfolio frame. The only generated content in this whole exercise lives in the clearly-labeled comparison image, presented as a teaching demonstration, never as the photograph.
⚙️ Settings Box — Dim window still life (a starting point, not a recipe)
Setting Start here Why Mode Aperture-priority (A/Av) or phone Pro You want to control depth of field and let the meter set time Aperture f/5.6–f/8 Enough depth for the whole object sharp; not so much you need even more light ISO Let it climb — ISO 3200–12800 We want noise here; this is the point of the exercise Shutter Whatever the meter gives (tripod/steady surface) Long is fine on a steady base; we're not freezing motion Focus Single-point AF on the near edge of the object The thing the photo is about must be sharp (Chapter 4) White balance Custom/Kelvin or "shade" The fading window goes cool/blue; decide the mood deliberately (Chapter 5) Support Tripod, books, or windowsill Long exposures in dim light demand a steady base RAW On Your negative, your authorship evidence, your latitude (Chapters 26, §33.5) Mobile-only: use Pro/RAW mode, brace the phone on the sill, tap the near edge to focus, and pull exposure down a touch so the shadow side stays moody. Your phone's "night mode" already runs AI denoise — turn it off for the first capture so you can see the raw noise, then compare.
The setup
Soft window light, one object, a dark interior to let it glow against shadow — the same controllable light you met in Chapter 13, now used as a clean test bench.
FIGURE CS33.2 — Kitchen-window still-life setup (top-down view)
[ WINDOW — fading dusk light, soft and cool ]
│
│ soft light rakes in from the left, low and weakening
▼
( S ) ── enamel mug, near edge toward camera
↙ ↘
○ (optional) [ dark room behind: no second light,
white card to so the mug glows against shadow ]
bounce a little
fill into the
shadow side
│
[ CAM ] on a tripod / braced on the counter, mug at its own level
Symbol legend (stated once): ( S ) = subject · ▢ = key light · ☀ = sun/available light ·
[CAM] = camera · ○ = reflector/fill · ▣ = rim/background light · arrows = light's direction toward subject.
Here the "key" is simply the window; the only choice is whether to bounce a little fill or let the shadow go.
We are using one soft source (the window) and an optional white card to lift the shadow side a stop or so — pure Chapter 13. The dim, late timing is the only twist, and it exists to manufacture the noise that the honest AI pipeline will then address.
The shoot, in phases
Phase 1 — Capture the real frame (noisy on purpose)
Turn the mug until the fading window light rakes across its near rim and reveals the chips and the texture of the enamel (Chapter 1's lesson: side light shapes; front light flattens). Frame it against the dark room so it glows. Let the ISO climb — you'll see the meter push to ISO 6400 or beyond as the light dies. Make a dozen frames, refining the turn of the mug and the crop. Keep the best.
FIGURE CS33.3 — "Enamel mug, last light" (the capture, before any edit) [constructed teaching example]
THE FRAME A chipped white enamel mug, near rim toward the lens, fills the right two-thirds of a
horizontal frame against a near-black interior. A thin rim of the dark room shows at left.
THE LIGHT Soft, cool, weakening window light from camera-left, low and raking; the near rim bright,
the far side falling two stops into shadow; a faint reflected highlight where the optional
white card lifts the shadow.
THE MOMENT No action — a still life — but timed to the last usable minute of window light, when it is
softest and coolest.
THE CHOICES f/5.6, ISO 6400, a slow shutter on a braced camera, single-point focus on the near rim;
framed to let the dark room set off the lit mug. RAW.
THE EFFECT The eye lands on the bright near rim, slides along the raking light into the chipped texture,
and rests in the shadow. It glows — but at 100% the shadow areas crawl with colored noise and
the fine enamel texture is half-buried in grain.
THE LESSON A genuinely good photograph can still carry a real technical problem (here, high-ISO noise)
that *honest* AI tools were built to solve — by refining the record, not changing it.
This is your raw material: a real, well-seen frame with a real flaw. Now we fix the flaw honestly.
Before we touch a slider, name the goal in one sentence, because it governs every choice that follows: we want the best possible version of what the light recorded — and nothing more. That sentence is the entire ethics of this shoot compressed to a rule. Every tool we are about to use either serves it (refines the record) or violates it (invents content). We will use only the first kind on the portfolio frame.
Phase 2 — AI denoise (refining the real signal)
Open the RAW. Apply AI denoise. Watch the colored speckle in the shadows melt away while the chipped texture stays — the denoiser has learned to tell noise from detail and strip the former (§33.2). Push it to its strongest setting and look at 100%: somewhere past "clean," it begins to smear — a fleck of glaze turns to plastic, a tiny scratch vanishes. That smear is the denoiser starting to invent a smoother surface than the light recorded. Back off to the point where it removes noise without erasing real texture. That point — where restoration would tip into invention — is the lesson of this phase. You are still firmly on the editing side: every pixel's source is the scene.
💡 Why It Works: AI denoise sits at the gentle end of the spectrum (Figure 33.2) because its job is to bring the image closer to the scene you photographed — the noise was a sensor artifact, not part of the world (Chapter 3). That is the same logic that made removing a sensor-dust spot enhancement in Chapter 29: you are subtracting something the camera added, not something the world contained. The instant the denoiser smooths away real texture, it stops restoring and starts fabricating — which is your cue to back off. Honest use is a dial, and you are the one who decides how far it turns.
Phase 3 — Mask and grade (local work on the real frame)
Use AI masking to select the mug separately from the background (Chapter 25's local adjustments, made effortless; Chapter 28's dodge-and-burn logic). Lift the near rim a touch, deepen the shadow side, warm the highlight just slightly against the cool ambient for a little tension (Chapter 27's grading). The mask invents nothing — it only understands which of your real pixels are mug and which are room, so your real adjustments land precisely. Crop and straighten (Chapter 26). The frame is now finished, and every pixel still traces to the light that reached the sensor.
It is worth pausing on why masking is the most ethically settled AI tool in the whole pipeline, because the reason is instructive. Every other tool we have touched does something to the pixels — denoise alters them, upscale adds them, grade shifts them. Masking touches no pixel at all. It performs a single act of understanding: "these pixels are mug; those are room." It is closer to a very fast intern who traces a selection for you than to any kind of image generation. Whatever you do after the mask — dodge, burn, warm, cool — is your own adjustment of your own real pixels, exactly as you would have done by hand with a brush, only faster and cleaner. That is why no serious objection has ever attached to it: there is no record to falsify when nothing was invented and nothing was even changed by the AI itself. Hold this as a clarifying case — when someone says "AI in photography is inherently dishonest," masking is the counterexample that proves the issue is never whether AI touched the file, but what it did.
A note for the mobile-only reader, because this entire pipeline is available in your pocket. Shoot the dim frame in your phone's Pro/RAW mode with night mode off so you capture honest noise. Then, in a capable mobile editor: the "denoise"/"clean" control is your Phase 2; the "select subject"/"select sky" tap is your Phase 3 mask; the tone, color, and crop tools finish it; the "enhance"/upscale button is your Phase 4 (used gently). The only tool to keep your finger off for the portfolio frame is the generative remove/"magic eraser" — that is the one that crosses to the generating side. Your phone can run the whole honest pipeline; the discipline, not the device, is what makes it honest.
Phase 4 — Modest upscale (only if you need the size)
Say you want a 13×19 print. Upscale 1.5×–2× with AI upscaling and inspect at 100%. At a modest factor the invented fine detail is invisible and harmless — it is plausibly extending texture the lens nearly resolved. Resist the temptation to push to 8×: that is where the upscaler starts inventing scratches and glaze patterns that were never there, and a careful eye (or a judge) will see them. Disclose if you pushed it hard; at a modest factor for a print, a note is courteous but the image is still fundamentally your capture.
Phase 5 — The comparison pair (the only generated content, openly labeled)
Now, as a teaching demonstration only, build the editing-vs-generating pair. Duplicate the finished frame. On the duplicate, use generative fill to remove the chips from the mug — invent smooth, un-chipped enamel where the real chips were. Place the two side by side: VERSION A (your honest edit, chips intact) and VERSION B (generative-filled, chips erased). A viewer cannot tell by eye that B's smooth surface is invented. Caption the pair plainly — "Left: real capture, honestly edited. Right: chips removed with generative fill — a fabrication, shown to teach the line." This labeled pair, not the generated image alone, is the artifact that has value: it teaches where editing ends and generating begins.
⚠️ Common Mistake: Letting the generated comparison image escape its frame — reposting Version B later without the label, "because it looks better." The whole exercise's integrity depends on Version B never circulating as a photograph. The fix is procedural: name the file
DEMO-generated-do-not-publish, keep it only inside the captioned pair, and treat the label as part of the image, not an optional caption. A fabrication is honest only while it is disclosed; the moment it travels without its label, you have made the exact false claim this chapter warns against.
Phase 6 — Read the provenance (the receiving side)
There is a second half to provenance you will need as a working photographer, and it is the mirror image of disclosure: not only attaching a record of how your image was made, but reading the record on an image someone hands you. A picture editor, a contest judge, a client, a collaborator — increasingly you will receive images whose history matters, and the skill is to ask for and interpret their provenance rather than trust your eye, which we have now proven cannot tell editing from generating.
Practice it on your own pair from Phase 5. Open both the honest frame and the chips-removed generation in software that can display content credentials, and look at what each one's record says — or fails to say. The honest frame's credentials should show a clean chain: captured by a camera, then a list of edits (denoise, masking, grade, crop) that are all record-refining. The generation's credentials, if the tool wrote them honestly, should reveal a generative step — content created or replaced by a model — which is exactly the flag a judge would look for. And if the credentials are missing entirely, that absence is itself information: an image presented as a captured photograph with no provenance and no RAW behind it deserves more scrutiny, not less.
FIGURE CS33.5 — Reading two images' provenance (what the record reveals)
HONEST FRAME (Version A) GENERATED-FILL FRAME (Version B)
┌──────────────────────────┐ ┌──────────────────────────┐
│ origin: captured (camera) │ │ origin: captured (camera) │
│ edits: │ │ edits: │
│ • AI denoise │ │ • AI denoise │
│ • masked local adjust │ │ • masked local adjust │
│ • grade, crop, upscale │ │ • grade, crop │
│ (all record-refining) │ │ • ►GENERATIVE FILL◄ │ ← the flag
│ RAW available: yes │ │ RAW available: yes │
└──────────────────────────┘ └──────────────────────────┘
chain is clean chain shows invention
No credentials + no RAW + presented as a photo = scrutinize, don't trust the eye.
The lesson of this phase is humbling and practical: in a world where the eye can no longer tell, the record — provenance you can read and verify — becomes the working photographer's real instrument of trust. Learn to attach it to your work and to demand it of others'.
🔗 Connection: This receiving-side skill is where Chapter 30's file discipline pays off again: the RAW you ingested, named, and backed up (the 3-2-1 rule) is the provenance you can produce on demand, and the metadata habits you built there are the foundation content credentials extend. Provenance is not a new chore; it is the old discipline of keeping your negatives and knowing your files, now made verifiable.
The edit pass and disclosure
Here is the finished honest frame and its disclosure — the deliverable.
FIGURE CS33.4 — "Enamel mug, last light" (finished, honest edit) [constructed teaching example]
THE FRAME Same composition as the capture: the chipped mug glowing against a near-black room, near rim
bright, chips intact and sharp.
THE LIGHT The recorded window light, now clean: shadows smooth instead of noisy, the cool ambient held
with a faint warm lift on the highlight for tension.
THE MOMENT The last minute of window light, preserved.
THE CHOICES AI-denoised (backed off before it smeared texture), AI-masked for local dodge/burn, graded,
cropped, modestly upscaled for print. NOTHING added or removed. RAW retained.
THE EFFECT It reads as exactly what it is — a real, well-seen still life, cleaned up honestly. The chips
are still there because they were there; the photograph is the best version of *what the
light recorded*.
THE LESSON An entire AI-assisted pipeline can run on the *editing* side of the line. Honest use is not
"no AI"; it is AI that refines the record plus a clear account of what was done.
The two-layer disclosure (§33.4): - Plain-language label (for humans): "Real capture at the kitchen window. AI denoise, AI-masked local adjustments, color grade, crop, modest upscale for print. No elements added or removed." - Content credentials (for verification): the embedded provenance should record: captured by camera (RAW), then edits — denoise, local masking adjustments, grade, crop, upscale — signed so a viewer's software can confirm the chain. Keep the original RAW as authorship evidence (§33.5, Chapter 30's backup).
Notice the disclosure is specific — it lists exactly what was done — because the more precise the label, the more it builds trust rather than merely satisfying a rule.
Discussion questions
- In Phase 2 you backed the denoiser off at the point where it started smearing real texture. In your own words, why is that the boundary between restoration and invention — and how does it echo the sensor-dust-vs.-real-person distinction from Chapter 29?
- The comparison pair (Phase 5) is the only place generated content appears, and it is openly labeled. Why is that honest, while quietly using the same generative-fill on the portfolio frame would not be?
- Your disclosure listed each edit specifically rather than just saying "AI was used." Argue why specific disclosure builds more trust than a vague one — and when a vague label might actually be worse than none.
- A client says they "don't care how it was made, just make the mug look perfect" and asks you to remove the chips for a product listing. Using §33.3 and §33.5, explain what you'd tell them and where you'd draw your line.
Your turn (extension)
Run the whole pipeline on a subject that matters to you — a person (with consent, Chapter 32), a meal you cooked, an object with a story — and produce three deliverables: (1) the finished honest frame, (2) its two-layer disclosure, and (3) the captioned editing-vs-generating comparison pair. Then write one paragraph answering the chapter's question about your honest frame: could a model have made this, and what does it have that a generation never could? File all of it in your Portfolio beside your written AI stance.
For a harder extension, run the pipeline twice on the same subject under two different lights — once at the bright midday window, once at the dying dusk window — and compare how much honest AICleanup each needs. The midday frame, shot at low ISO, will barely need denoise; the dusk frame will lean on it heavily. Seeing that the honest tools work hardest exactly where your capture was weakest is a quiet but important lesson: AI on the editing side is a way to recover what difficult light cost you, not a substitute for seeing the light in the first place (Chapter 5). The photographer who nails the light needs the least AI — which tells you where to put your effort.
Key takeaways
- An entire AI-assisted edit — denoise, masking, grade, crop, modest upscale — can run honestly on the editing side of the line, because every operation refines what the light recorded rather than inventing new content. Honest use is not "no AI"; it is AI that refines plus a clear account of what was done.
- The boundary inside a single tool is real and findable: AI denoise restores until the point it starts smearing real texture, where it begins to invent — and you, not the slider, decide where to stop.
- The disclosed comparison pair is the artifact with teaching value, not the generated image alone; a fabrication is honest only while it stays labeled and never circulates as a photograph.
- Disclose on two layers and be specific: a plain-language sentence listing each edit, plus embedded content credentials, with the RAW kept as authorship and provenance evidence. Precise disclosure builds trust; vague disclosure can erode it.