marketing-photography
Photography for a marketing site or landing page, generated with an AI image
model. Two constraints make this a different job from ad creative: the same
people must recur across many pages and sections, and each image must earn its
place next to specific copy.
ad-creative-generation covers ad creatives: single images, style references,
disposable. Each render stands alone, so a new face every time costs nothing.
On a marketing site it costs everything. A firm whose people change on every
page reads as stock photography, and visitors stop trusting the page. This
skill fixes identity first, then matches every shot to the section it sits in.
For starting prompt recipes see ad-image-prompt-library. For text-heavy or
UI-heavy creatives rendered as HTML see ad-creative-templates.
When to use
- Triggers: 'site photography', 'consistent people across pages', 'landing page images', 'photoreal brand imagery', 'AI photos that do not look AI'.
- A written cast with one 4-frame casting sheet per character, every scene an edit against those sheets (a full animation-style model sheet hurts photoreal identity), and each shot matched to its section's claim.
- Realism levers (real camera and lens, motivated mixed light, unretouched skin, mundane objects, nobody facing the camera), designing garbled text out, sector-cliche negatives, honesty rules (no names, testimonials or badges), and QA with pre-sized AVIF and WebP. Generic CHARACTER and SCENE templates; pairs with ad-image-prompt-library.
1. Use this skill when it owns the job
| You need |
Reach for |
| Consistent people across site pages and sections |
This skill |
| One disposable image for an ad |
ad-creative-generation |
| A starting prompt or format recipe |
ad-image-prompt-library |
| Pixel-exact text, UI, or logo inside the image |
ad-creative-templates (an HTML render, not a model) |
| Photos of real staff or customers |
A camera. Never generate stand-ins for real people (see §8) |
2. Build a cast before you generate any scene
The failure mode: generating each scene from scratch yields a different face
every time. A site where nobody recurs reads as stock. Fix identity before you
shoot anything.
Define each character as a written description first. Specific and unglamorous
beats beautiful: small flaws lock identity better than polished features. Each
description covers:
- Age and build, stated plainly.
- Hair, including how it is untidy: a cowlick, grown-out roots, a frayed cut.
- Skin, including asymmetries and blemishes: uneven tone, a mole, lines, shine.
- Wardrobe, including wear and creases: a faded cuff, a stretched collar,
scuffed shoes.
- One or two identity anchors: a scar, a pair of glasses, a lanyard, a watch.
Anchors are what the model holds onto across scenes.
Then generate one casting reference sheet per character: four frames in a 2x2
grid on a seamless mid-grey background. Front head-and-shoulders,
three-quarter view, strict profile, full body. Identical wardrobe, hair, and
light in all four frames. Neutral expression throughout.
Then generate every scene as an edit call that passes the relevant sheet or
sheets as reference images, with an identity-lock line: keep the exact same
face, features, and proportions as the reference, and do not restyle, age,
slim, or beautify. Identity carries through the reference, not through
re-describing the person each time.
Store the sheets at durable URLs and commit the cast definitions to the repo.
A cast that lives only in a prompt thread is lost, and page two gets different
faces.
3. Why not a full animation-style model sheet
The trap: the popular character design reference sheet prompt (turnarounds
plus an expression grid plus material studies plus a colour palette plus
callouts) is built for animation production, where a team of artists reads the
board. Fed back into a photoreal edit model as a reference image, a busy
9-panel board degrades identity transfer. The model tries to reproduce the
board layout instead of the person. Expressions leak into scenes, palette
strips show up as props, callout lines sharpen into artifacts.
For photoreal work, strip the board to the clean 4-frame casting card. Keep
the identity-lock discipline, drop the production-board furniture.
| Animation board |
Casting card (this skill) |
| 9 or more panels: turnarounds, expressions, materials, palette, callouts |
4 frames: front, three-quarter, profile, full body |
| Made for artists to read |
Made for an edit model to match |
| Expression grid |
Neutral expression in every frame |
| Palette strips and labels |
No text, no swatches, no graphics |
4. Match the image to the section, not to the page
Derive the shot list from copy that already exists. Read the page first. Each
image answers one claim the section makes. If you cannot name the sentence an
image supports, do not generate it.
| Section type |
What to shoot |
| How it works |
The work being done: hands, tools, and materials mid-task |
| Audience or segment cards |
One professional per segment, shot as a consistent series with matched framing and grade |
| Human review or trust |
The review happening: a person reading, checking, annotating, discussing |
| Stats or numbers |
The thing being counted, in its real setting |
Rhythm matters as much as subject. Do not put a photo in every section.
Alternate image-led sections with dense text sections so each photo keeps its
weight. Respect the tonal bands of the design: when the page alternates light
and dark bands, place and grade images to sit inside that rhythm, not against
it.
5. The realism levers
What stops a render reading as AI:
- Name a real camera body, a real lens, and a working aperture: a full-frame
mirrorless body, a 35mm f/1.8 prime, shot at f/2.8. Never write "8k",
"hyperrealistic", "stunning", or "masterpiece". Those words push toward the
glossy AI look.
- Light the scene from named, motivated sources, and let them mix and
conflict: overcast daylight from a window at left plus warm overhead
fluorescents, with a green cast where they meet. Unflattering light reads as
real.
- Nobody looks at the camera. Nobody smiles for the camera. No handshakes, no
pointing at charts, no arms-crossed power poses. People in real rooms are
busy with something else.
- Ask for unretouched skin: pores, lines, uneven tone, stray hairs, natural
asymmetry, a little shine.
- Dress the set with specific mundane objects: a dented travel mug, a ring
mark on the desk, loose staples, a cardigan over a chair back, a slightly
crooked wall frame.
- Frame it as a real photographer would. Name the vantage point ("shot from
the doorway, slightly below eye level") and let the horizon sit slightly off
level.
6. Text is the giveaway: design it out
Image models garble text. On a document-heavy page that produces convincing
nonsense, which is worse than no image at all. A page of blurred fake
paragraphs next to real copy destroys the realism every other lever built. So
prevent text instead of fixing it:
- Angle documents so type reads as grey texture. Turn pages away, shoot papers
from above at an angle, hold shallow depth of field on the text.
- Angle screens away from the camera and keep them out of focus.
- Specify book spines, whiteboards, and signage with no lettering.
- Put "no text, no lettering, no numbers, no labels, no captions, no
watermark, no logo" in every negative list.
- Then zoom in at full resolution during QA and check. Thumbnails hide garbled
type. Full resolution shows it.
7. Negative prompts as a cliche list
A cliche is what makes a page look cheap, and every sector has its own
saturated stock imagery. Research yours first, then convert it into a negative
list.
The method: open three to five competitor or peer sites and look at the images
they actually render, then read the metadata behind them. Filenames and alt text
leak the stock library and the subject ("diverse-team-meeting",
"handshake-office", "iStock-504309690"), so they are strong supporting evidence.
They are not sufficient on their own: a site can carry its entire hero set as
CSS backgrounds and expose no <img> tag at all, and lazy-loaded images and
empty alt attributes hide the rest. Scroll the page, let the images load, and
judge what you can see; use filenames and alt text to confirm what you are
looking at. The repeats are the sector cliches. Negate them by name.
A generic starter list, extended with whatever the research finds:
no handshake, no pointing at a screen, no arms crossed, no headset smile,
no staged high-five, no puzzle pieces, no chess pieces, no rocket launch,
no magnifying glass over a chart, no text, no lettering, no numbers,
no labels, no captions, no watermark, no logo, no plastic skin,
no studio polish, no oversaturation
8. Honesty rules for generated people
Generated people are not real staff and not real customers. Treat that line as
a hard boundary, not as tone guidance.
- Never attach a name, title, credential, or quotation to a generated face. No
testimonials with generated portraits. No "meet our team". No staff roster.
- Alt text describes the scene, never the individual. An occupational role is
fine when the visible action shows it: "A reviewer annotating printed pages at
a shared table" is accurate and useful to a screen reader. What is banned is
anything that identifies a particular person, real or implied: a name, a
title held by someone, a credential, or a claim of employment. Write "A
reviewer annotating printed pages", never "Dr. Alvarez, our lead reviewer".
- Never generate an accreditation seal, certification badge, or compliance
mark into an image. A photographed badge is a fabricated claim.
- Never reuse one placeholder headshot across several differently named
people. That is a real and common failure, and readers spot it fast. It
destroys credibility in a way no other defect does.
- This is exactly why the cast stays unnamed. The cast exists so the same
faces recur honestly as unnamed illustration, never as invented staff.
9. QA loop and production notes
- Judge the casting sheet before generating any scene. One bad sheet poisons
every scene that uses it, so a flaw you accept here costs you the whole set.
- Repair drift at the level that caused it, and do not confuse the two cases. If
ONE scene drifts off an otherwise good sheet, regenerate that scene: the sheet
is still the approved identity and replacing it would invalidate every scene
already signed off. Only regenerate the character when the SHEET itself is
wrong, and then regenerate every scene built on it.
- Review the audience-card series together as a grid. Brightness and grade
inconsistency only shows side by side. Single-image review passes frames
that clash on the page.
- Serve pre-sized AVIF plus WebP. Derive the declared width and height from
the encoded files, not from the arithmetic: a resampler can land a pixel off
the intended size, and a declared size that misses by one pixel causes layout
shift.
- Keep costs concrete but model-agnostic: budget roughly $0.05-0.15 per image
at current reference-model pricing. Price the run before batching: cast
sheets plus scenes plus retries, times the per-image rate.
10. Copy-paste templates
Fill every [PLACEHOLDER]. Keep the order throughout: subject, what is
happening, the room, light, camera, NOT WANTED. Readers copy these directly.
CHARACTER template. One per cast member, committed to the repo:
[CAST ID, e.g. CAST-01. Never a personal name.]
Subject: [AGE], [BUILD], [HAIR including how it is untidy],
[SKIN including asymmetries and blemishes], wardrobe:
[GARMENTS including wear and creases].
Identity anchors: [ANCHOR 1], [ANCHOR 2].
Casting sheet: 2x2 grid on a seamless mid-grey background.
Frame 1: front head-and-shoulders, neutral expression.
Frame 2: three-quarter view, neutral expression.
Frame 3: strict profile, neutral expression.
Frame 4: full body, standing, neutral expression.
Identical wardrobe, hair, and light in all four frames.
NOT WANTED: text, labels, graphics, palette strips, expression
grid, extra panels, retouched skin, glamour lighting.
SCENE template:
Subject: [CAST ID(s) from the casting sheet(s), passed as
reference images]. Keep the exact same face, features, and
proportions as the reference. Do not restyle, age, slim, or
beautify.
What is happening: [ONE ACTION that answers the section claim.
Nobody looks at the camera. Nobody smiles for the camera.]
The room: [SETTING, e.g. a clinic-style setting]. Dressing:
[MUNDANE OBJECT 1], [MUNDANE OBJECT 2].
Documents and screens: [ANGLED AWAY / OUT OF FOCUS /
NO LETTERING, as applicable].
Light: [SOURCE 1 with direction] plus [SOURCE 2 with
direction]. Let them mix: [WHERE THEY MEET, e.g. the green
cast along one wall].
Camera: [BODY], [LENS] at [APERTURE], [VANTAGE POINT, e.g.
from the doorway, slightly below eye level]. Horizon
slightly off level.
NOT WANTED: no text, no lettering, no numbers, no labels, no
captions, no watermark, no logo, no [SECTOR CLICHES from §7],
no plastic skin, no studio polish, no oversaturation.
11. Checklist: one photography pass
Pairs with: ad-creative-generation (disposable single ad images, style
references, the vision-QA loop), ad-image-prompt-library (starting prompt
recipes and realism levers), ad-creative-templates (anything with exact text,
UI, or logo).
1---2name: marketing-photography3description: "Use when a marketing site or landing page needs believable AI photography where the same people recur across sections and each image matches the copy beside it. For one-off ad images use ad-creative-generation."4metadata:5 source: https://skillmd.com/skills/pooriaarab/marketing-photography6---78# marketing-photography910Photography for a marketing site or landing page, generated with an AI image11model. Two constraints make this a different job from ad creative: the same12people must recur across many pages and sections, and each image must earn its13place next to specific copy.1415`ad-creative-generation` covers ad creatives: single images, style references,16disposable. Each render stands alone, so a new face every time costs nothing.17On a marketing site it costs everything. A firm whose people change on every18page reads as stock photography, and visitors stop trusting the page. This19skill fixes identity first, then matches every shot to the section it sits in.20For starting prompt recipes see `ad-image-prompt-library`. For text-heavy or21UI-heavy creatives rendered as HTML see `ad-creative-templates`.2223## When to use2425- **Triggers:** 'site photography', 'consistent people across pages', 'landing page images', 'photoreal brand imagery', 'AI photos that do not look AI'.26- A written cast with one 4-frame casting sheet per character, every scene an edit against those sheets (a full animation-style model sheet hurts photoreal identity), and each shot matched to its section's claim.27- Realism levers (real camera and lens, motivated mixed light, unretouched skin, mundane objects, nobody facing the camera), designing garbled text out, sector-cliche negatives, honesty rules (no names, testimonials or badges), and QA with pre-sized AVIF and WebP. Generic CHARACTER and SCENE templates; pairs with ad-image-prompt-library.2829## 1. Use this skill when it owns the job3031| You need | Reach for |32|---|---|33| Consistent people across site pages and sections | This skill |34| One disposable image for an ad | `ad-creative-generation` |35| A starting prompt or format recipe | `ad-image-prompt-library` |36| Pixel-exact text, UI, or logo inside the image | `ad-creative-templates` (an HTML render, not a model) |37| Photos of real staff or customers | A camera. Never generate stand-ins for real people (see §8) |3839## 2. Build a cast before you generate any scene4041The failure mode: generating each scene from scratch yields a different face42every time. A site where nobody recurs reads as stock. Fix identity before you43shoot anything.4445Define each character as a written description first. Specific and unglamorous46beats beautiful: small flaws lock identity better than polished features. Each47description covers:4849- Age and build, stated plainly.50- Hair, including how it is untidy: a cowlick, grown-out roots, a frayed cut.51- Skin, including asymmetries and blemishes: uneven tone, a mole, lines, shine.52- Wardrobe, including wear and creases: a faded cuff, a stretched collar,53 scuffed shoes.54- One or two identity anchors: a scar, a pair of glasses, a lanyard, a watch.55 Anchors are what the model holds onto across scenes.5657Then generate one casting reference sheet per character: four frames in a 2x258grid on a seamless mid-grey background. Front head-and-shoulders,59three-quarter view, strict profile, full body. Identical wardrobe, hair, and60light in all four frames. Neutral expression throughout.6162Then generate every scene as an edit call that passes the relevant sheet or63sheets as reference images, with an identity-lock line: keep the exact same64face, features, and proportions as the reference, and do not restyle, age,65slim, or beautify. Identity carries through the reference, not through66re-describing the person each time.6768Store the sheets at durable URLs and commit the cast definitions to the repo.69A cast that lives only in a prompt thread is lost, and page two gets different70faces.7172## 3. Why not a full animation-style model sheet7374The trap: the popular character design reference sheet prompt (turnarounds75plus an expression grid plus material studies plus a colour palette plus76callouts) is built for animation production, where a team of artists reads the77board. Fed back into a photoreal edit model as a reference image, a busy789-panel board degrades identity transfer. The model tries to reproduce the79board layout instead of the person. Expressions leak into scenes, palette80strips show up as props, callout lines sharpen into artifacts.8182For photoreal work, strip the board to the clean 4-frame casting card. Keep83the identity-lock discipline, drop the production-board furniture.8485| Animation board | Casting card (this skill) |86|---|---|87| 9 or more panels: turnarounds, expressions, materials, palette, callouts | 4 frames: front, three-quarter, profile, full body |88| Made for artists to read | Made for an edit model to match |89| Expression grid | Neutral expression in every frame |90| Palette strips and labels | No text, no swatches, no graphics |9192## 4. Match the image to the section, not to the page9394Derive the shot list from copy that already exists. Read the page first. Each95image answers one claim the section makes. If you cannot name the sentence an96image supports, do not generate it.9798| Section type | What to shoot |99|---|---|100| How it works | The work being done: hands, tools, and materials mid-task |101| Audience or segment cards | One professional per segment, shot as a consistent series with matched framing and grade |102| Human review or trust | The review happening: a person reading, checking, annotating, discussing |103| Stats or numbers | The thing being counted, in its real setting |104105Rhythm matters as much as subject. Do not put a photo in every section.106Alternate image-led sections with dense text sections so each photo keeps its107weight. Respect the tonal bands of the design: when the page alternates light108and dark bands, place and grade images to sit inside that rhythm, not against109it.110111## 5. The realism levers112113What stops a render reading as AI:114115- Name a real camera body, a real lens, and a working aperture: a full-frame116 mirrorless body, a 35mm f/1.8 prime, shot at f/2.8. Never write "8k",117 "hyperrealistic", "stunning", or "masterpiece". Those words push toward the118 glossy AI look.119- Light the scene from named, motivated sources, and let them mix and120 conflict: overcast daylight from a window at left plus warm overhead121 fluorescents, with a green cast where they meet. Unflattering light reads as122 real.123- Nobody looks at the camera. Nobody smiles for the camera. No handshakes, no124 pointing at charts, no arms-crossed power poses. People in real rooms are125 busy with something else.126- Ask for unretouched skin: pores, lines, uneven tone, stray hairs, natural127 asymmetry, a little shine.128- Dress the set with specific mundane objects: a dented travel mug, a ring129 mark on the desk, loose staples, a cardigan over a chair back, a slightly130 crooked wall frame.131- Frame it as a real photographer would. Name the vantage point ("shot from132 the doorway, slightly below eye level") and let the horizon sit slightly off133 level.134135## 6. Text is the giveaway: design it out136137Image models garble text. On a document-heavy page that produces convincing138nonsense, which is worse than no image at all. A page of blurred fake139paragraphs next to real copy destroys the realism every other lever built. So140prevent text instead of fixing it:141142- Angle documents so type reads as grey texture. Turn pages away, shoot papers143 from above at an angle, hold shallow depth of field on the text.144- Angle screens away from the camera and keep them out of focus.145- Specify book spines, whiteboards, and signage with no lettering.146- Put "no text, no lettering, no numbers, no labels, no captions, no147 watermark, no logo" in every negative list.148- Then zoom in at full resolution during QA and check. Thumbnails hide garbled149 type. Full resolution shows it.150151## 7. Negative prompts as a cliche list152153A cliche is what makes a page look cheap, and every sector has its own154saturated stock imagery. Research yours first, then convert it into a negative155list.156157The method: open three to five competitor or peer sites and look at the images158they actually render, then read the metadata behind them. Filenames and alt text159leak the stock library and the subject ("diverse-team-meeting",160"handshake-office", "iStock-504309690"), so they are strong supporting evidence.161They are not sufficient on their own: a site can carry its entire hero set as162CSS backgrounds and expose no `<img>` tag at all, and lazy-loaded images and163empty alt attributes hide the rest. Scroll the page, let the images load, and164judge what you can see; use filenames and alt text to confirm what you are165looking at. The repeats are the sector cliches. Negate them by name.166167A generic starter list, extended with whatever the research finds:168169```text170no handshake, no pointing at a screen, no arms crossed, no headset smile,171no staged high-five, no puzzle pieces, no chess pieces, no rocket launch,172no magnifying glass over a chart, no text, no lettering, no numbers,173no labels, no captions, no watermark, no logo, no plastic skin,174no studio polish, no oversaturation175```176177## 8. Honesty rules for generated people178179Generated people are not real staff and not real customers. Treat that line as180a hard boundary, not as tone guidance.181182- Never attach a name, title, credential, or quotation to a generated face. No183 testimonials with generated portraits. No "meet our team". No staff roster.184- Alt text describes the scene, never the individual. An occupational role is185 fine when the visible action shows it: "A reviewer annotating printed pages at186 a shared table" is accurate and useful to a screen reader. What is banned is187 anything that identifies a particular person, real or implied: a name, a188 title held by someone, a credential, or a claim of employment. Write "A189 reviewer annotating printed pages", never "Dr. Alvarez, our lead reviewer".190- Never generate an accreditation seal, certification badge, or compliance191 mark into an image. A photographed badge is a fabricated claim.192- Never reuse one placeholder headshot across several differently named193 people. That is a real and common failure, and readers spot it fast. It194 destroys credibility in a way no other defect does.195- This is exactly why the cast stays unnamed. The cast exists so the same196 faces recur honestly as unnamed illustration, never as invented staff.197198## 9. QA loop and production notes199200- Judge the casting sheet before generating any scene. One bad sheet poisons201 every scene that uses it, so a flaw you accept here costs you the whole set.202- Repair drift at the level that caused it, and do not confuse the two cases. If203 ONE scene drifts off an otherwise good sheet, regenerate that scene: the sheet204 is still the approved identity and replacing it would invalidate every scene205 already signed off. Only regenerate the character when the SHEET itself is206 wrong, and then regenerate every scene built on it.207- Review the audience-card series together as a grid. Brightness and grade208 inconsistency only shows side by side. Single-image review passes frames209 that clash on the page.210- Serve pre-sized AVIF plus WebP. Derive the declared width and height from211 the encoded files, not from the arithmetic: a resampler can land a pixel off212 the intended size, and a declared size that misses by one pixel causes layout213 shift.214- Keep costs concrete but model-agnostic: budget roughly $0.05-0.15 per image215 at current reference-model pricing. Price the run before batching: cast216 sheets plus scenes plus retries, times the per-image rate.217218## 10. Copy-paste templates219220Fill every `[PLACEHOLDER]`. Keep the order throughout: subject, what is221happening, the room, light, camera, NOT WANTED. Readers copy these directly.222223CHARACTER template. One per cast member, committed to the repo:224225```text226[CAST ID, e.g. CAST-01. Never a personal name.]227228Subject: [AGE], [BUILD], [HAIR including how it is untidy],229[SKIN including asymmetries and blemishes], wardrobe:230[GARMENTS including wear and creases].231Identity anchors: [ANCHOR 1], [ANCHOR 2].232233Casting sheet: 2x2 grid on a seamless mid-grey background.234Frame 1: front head-and-shoulders, neutral expression.235Frame 2: three-quarter view, neutral expression.236Frame 3: strict profile, neutral expression.237Frame 4: full body, standing, neutral expression.238Identical wardrobe, hair, and light in all four frames.239240NOT WANTED: text, labels, graphics, palette strips, expression241grid, extra panels, retouched skin, glamour lighting.242```243244SCENE template:245246```text247Subject: [CAST ID(s) from the casting sheet(s), passed as248reference images]. Keep the exact same face, features, and249proportions as the reference. Do not restyle, age, slim, or250beautify.251252What is happening: [ONE ACTION that answers the section claim.253Nobody looks at the camera. Nobody smiles for the camera.]254255The room: [SETTING, e.g. a clinic-style setting]. Dressing:256[MUNDANE OBJECT 1], [MUNDANE OBJECT 2].257Documents and screens: [ANGLED AWAY / OUT OF FOCUS /258NO LETTERING, as applicable].259260Light: [SOURCE 1 with direction] plus [SOURCE 2 with261direction]. Let them mix: [WHERE THEY MEET, e.g. the green262cast along one wall].263264Camera: [BODY], [LENS] at [APERTURE], [VANTAGE POINT, e.g.265from the doorway, slightly below eye level]. Horizon266slightly off level.267268NOT WANTED: no text, no lettering, no numbers, no labels, no269captions, no watermark, no logo, no [SECTOR CLICHES from §7],270no plastic skin, no studio polish, no oversaturation.271```272273## 11. Checklist: one photography pass274275- [ ] Shot list derived from existing copy; every image names its sentence.276- [ ] Cast defined in writing; sheets generated; sheets judged before any scene.277- [ ] Sheets stored at durable URLs; cast definitions committed.278- [ ] Scenes built as edits against the sheets, with identity-lock lines.279- [ ] Real camera, lens, and aperture named; banned adjectives absent.280- [ ] Text designed out; full-resolution zoom check done.281- [ ] Sector cliches researched; negatives extended.282- [ ] Honesty rules hold: no names, no testimonials, no badges, scene-only alt283 text.284- [ ] Series reviewed as a grid; AVIF plus WebP pre-sized from encoded files.285286Pairs with: `ad-creative-generation` (disposable single ad images, style287references, the vision-QA loop), `ad-image-prompt-library` (starting prompt288recipes and realism levers), `ad-creative-templates` (anything with exact text,289UI, or logo).