A real personal
branding shoot.
At $0.04 a photo.

Before we start

I needed photos. Not a thousand, just 5 or 6 good photos of me, usable on LinkedIn, my site, as a newsletter hero, on a sales page. You know the problem: the last real photo session you did was years ago, your selfies are from 2022, and everything you post looks a little washed out.

Two classic options · pay a photographer (around €500 a session in Lyon or Paris for a decent set), or take out a subscription to an off-the-shelf generator like Klayn.ai (about €200/month). I found both too expensive for a need that comes up once a quarter.

So I did what I do every time I find a product too expensive · I looked under the hood. Klayn really just stacks three or four AI image models available on platforms like fal.ai. Per unit, each photo costs between $0.04 and $0.20 · not $200 a month. The rest is just interface, marketing and margin.

So I built my own tool. Here's the photo I'm proudest of (generated from 4 selfies shot on my iPhone):

How it got started (and the idea isn't mine)

To be honest, the idea isn't mine. It was Florian, our in-house videographer-photographer, who kicked the whole thing off.

A few months ago, Florian built a tool that generates product photos for e-commerce sites. The idea is dead simple · instead of organizing a shoot (gear, studio, photographer, post-production · easily €1,000 for a handful of products), you upload a photo of your product, describe the scene you want ("our cream in close-up on a black marble counter, golden light"), and the AI hands you 5-10 pro photos of your product in that scene. For a few cents each.

When I saw his first results on products, something clicked · if it works for products, why not for faces?

Because we all have the same problem · a LinkedIn profile photo from 2022, no portrait for the site's "about" page, nothing usable as a newsletter hero, no visual for sales pages. And between the €500 photo session and the €200/month Klayn subscription, there's a gap nobody fills. Especially for teams · if you have 5 people who each need 10 photos, it costs you an arm and a leg the classic way (the logistics of getting a photographer to the office, a whole day blocked off, a €2,000-3,000 final bill).

So I took Florian's same architecture, forked it, and rethought the whole thing for faces · handling identity references, prompts specific enough that people actually recognize the person, a "natural iPhone" mode by default to avoid the fake-magazine look.

Two weeks later, I had my photos. And then we said "hey, we should try it for other people at the company" · you'll get a glimpse a bit further down in the article.

What the result looks like

I built a dead-simple website. Four steps, in order:

In the dashboard, you also get a cost counter in the top right · today $X.XX, this month $Y.YY. So you don't burn through €30 without noticing (more on that below, because that's exactly what happened to me at the start).

Here are my 4 reference selfies

So you can see just how ridiculously simple step 01 is · here are exactly the 4 photos I uploaded into the tool to generate every photo you see in this article. 4 iPhone selfies, a light-colored wall, shot in 2 minutes:

That's it. No gear, no session, no suit. That is enough for the tool to understand my face and reproduce it in 50 different settings. The talent is in the prompt, not in the input photo.

The app, for real

Here's what it looks like once deployed. No terminal, no YAML, just a UI. You click, you paste in your 3 keys, you generate your photos.

The wizard you'll have on your end · home, example showcase, the "Start" button that kicks off the 4 steps (selfies → scene → proposals → renders). It's the page you see first when you land on your Vercel URL.

The Settings page that tells you in green that everything works · you get here from anywhere via the topbar. A "how it works" diagram, 3 cards for the 3 keys (fal.ai, OpenRouter, Admin token), a "Test connection" button under each field. When everything is green, you're ready.

Close-up on the cards · on the left the status (✓ Configured with the value masked as ••••XXXX, or ✗ Missing), on the right the "Test connection" button that pings fal.ai's billing endpoint. You expand "How to get it" if you're looking for the docs to follow.

And if you want to run it locally · same page, further down. Step-by-step instructions to clone the repo, run npm install, create your .env.local. It's the only moment in the tool where anyone mentions an env file · otherwise everything goes through the UI.

The photos that actually work

Let me show you two results · it gives you an idea of what you get when everything lines up.

Why this one works · the AI kept my real face (skull shape, beard structure, gaze), the light is soft and natural (no harsh flash), and black-and-white hides the sometimes too-smooth AI skin. It's the kind of portrait I can use as a profile photo without anyone asking questions.

Why this one works · the setting (rooftop at sunset) stays plausible, I'm holding the phone in a natural way, and the golden light masks the imperfections the AI sometimes puts on skin. Perfect as a sales-page or newsletter hero.

Gallery · what you can actually produce

So you don't stop at 2 photos · here's what the tool produces across 8 typical use cases for solid personal branding. Each time I give you the target use, the real cost, and why it works (because understanding why is what lets you replicate it on your end).

Wall of results · the raw volume, no commentary

Before we go through the cases one by one, just look at the contact sheet. 9 results side by side, like a photo studio handing you the sheet of thumbnails after the session. This is what an hour of generation at $0.04 a render looks like:

Now let's run through the cases in detail.

1 · Corporate LinkedIn (profile photo + About page)

• For what · LinkedIn profile photo, the site's "about" page, email signature, pitch intro slide.
• Cost · $0.04 a render, I took 6 before landing this one (≈ $0.25).
• What works · an "open space premium" mood but with natural depth blur (no stock bokeh), a controlled but not frozen expression, a slightly off-center gaze (not dead-center = it breathes). You can put it on LinkedIn tomorrow morning and nobody asks a thing.

2 · Speaker / stage (talk announcement, conference recap post, bio slide)

• For what · announcing an upcoming talk (meetup, industry conference, video podcast), illustrating a "takeaways after the conference" post, a bio slide in a client presentation.
• Cost · $0.04. A single generation, it was good on the first try.
• What works · the static stance near the lectern, the cold projector light behind (the signature of conference rooms), the well-fitted navy polo = the visual code of "expert taking the mic." You can use it as the hero of an "I give talks" page and nobody will go check.

3 · Café lifestyle (LinkedIn carousel, newsletter background)

• For what · a "my morning routine" carousel, a newsletter header, a course sales-page visual, a blog article illustration about productivity.
• Cost · $0.04, first render.
• What works · it's a close-up with no face · the AI doesn't have to fight to reproduce my face, so the result is ultra-clean (the AI's perfect skin only causes trouble on frontal portraits). Cinematic framing, golden light, a bookshelf that suggests "I read, therefore I think." You can spin this aesthetic across 50 visuals at once.

4 · Premium office tower (About hero, team page, pro photo)

• For what · the hero of the site's "About" page, an official photo for a press release, the opening slide of an investor pitch, a startup's team-page header.
• Cost · $0.04, I generated 3 to get a clean angle (≈ $0.12).
• What works · the high floor with a blurred skyline behind that suggests "a company that scales" (a premium office = a success signal), the hands-in-pockets stance that balances the understatement of the beige bomber without falling into the three-piece suit, the side natural light that masks the skin. A photo you keep for 2 years as an "about" hero without replacing it.

5 · Rooftop hero (homepage hero, pre-order page, newsletter hero)

• For what · a homepage hero, a pre-order page header, the opening photo of a product launch.
• Cost · $0.04. Count on 3-4 tries to find the clean angle (≈ $0.15).
• What works · the golden sunset light naturally masks the imperfections the AI puts on skin · that's the secret. A natural pose (phone in hand, gaze toward the horizon), an "end of a business day but chill" mood. Put this photo as your hero and you've got a site that looks seed-funded.

6 · Timeless B&W portrait ("main" profile photo, interview, press)

• For what · your main profile photo (LinkedIn, Twitter, site, guest podcast), a press photo to send a journalist, an interview cover.
• Cost · $0.04. It's my favorite, I generated about ten to find it (≈ $0.40).
• What works · black-and-white hides the AI's sometimes too-smooth skin (that's the main tell you spot on poorly dialed-in AI photos · skin that's too clean). Turtleneck = timeless, direct gaze into the camera = authority, soft side light = no harsh flash cutoff. This is the portrait you can leave up for 3 years without replacing it.

7 · Library café variant (podcast cover, "about" header)

• For what · a personal podcast cover, an "about" page header in a warmer mode, an author photo on a personal blog, an illustration for an essay or a book.
• Cost · $0.04, 2 renders to get the right angle (≈ $0.08).
• What works · a warmer mood than the corporate one (#1), more composed than the office tower (#4). Wood, tungsten light, a slightly blurred bookshelf background · it's the "accessible intellectual" portrait. Perfect when you sell training, consulting, written content. Very different from the corporate LinkedIn one · same face, two distinct visual universes.

8 · Morning home office (casual LinkedIn post, personal blog photo, course hero)

• For what · a "here's what my morning looks like" LinkedIn post, an illustration photo on a personal blog (productivity, routine, entrepreneur life), the hero of an online course sales page, an essay cover header.
• Cost · $0.04. 2 renders to get the mug at the right height (≈ $0.08).
• What works · the camel bomber + white tee = the "casual and owning it" signal without dropping into a hoodie, the mug in hand at the window = a time anchor for "it's morning, calm," the bookshelf + frames behind that whisper without shouting "I've got a world of my own." It's the photo that says "I'm expensive but I've got time" without having to say it.

9 · Before / after · the iPhone selfie vs what the AI turns it into

So you can gauge the raw delta, here side by side is an input iPhone selfie and an AI result generated from it:

That's exactly the jump · on the left, what's already in your phone's camera roll, on the right, what you can put on LinkedIn Monday morning. No haircut change, no shave, no studio. Just 4 cents and 30 seconds.

And it works for teammates too

That's where the tool gets really interesting. Once it was dialed in for my face, we could reuse it as-is for everyone else's faces at the company.

Here's a portrait generated for one of our teammates · 4 selfies sent as input, the brief "neutral black-and-white studio, direct gaze into the camera, understated suit," and a few renders later we had his full set of pro photos:

What this changes in practice · if you lead a team, you no longer need to bring a photographer to the office for everyone (a logistics nightmare, a price that climbs fast · between €1,500 and €3,000 for 5-10 people depending on the city). Each person sends their 4 selfies from home, you run the tool for 5 minutes, and everyone has their set. Once the tool is set up, the marginal cost of an extra teammate is around €0.80-1.

It's also the moment when the argument "I buy the code once instead of subscribing" makes complete sense · across 5-10 people to kit out, the software's ROI is immediate.

The shamefully failed photos (because you have to be honest)

I'm not going to sell you a dream. For every 1 photo that works, I throw out 4. Here are 3 characteristic flops · they give you an idea of the kind of traps to avoid.

Flop 1 · The AI invented a character who isn't me

That guy isn't me. I'm bald, I don't wear earrings, I've never had that haircut. Why it happens · when the prompt is too vague (something like "business man on a rooftop"), the model invents the character. My 4 reference selfies weigh less than the thousands of corporate LinkedIn photos the model saw during training.

The fix that worked · forcing very precise identity anchors into the prompt · "bald, salt-and-pepper beard, square face structure, blue eyes." And running a mini-test of 1 photo before the batch, to see if the AI is really catching my face. If it confuses me with someone else, I adjust before generating 5 photos at once.

Flop 2 · The frozen "2018 corporate stock photo" pose

This one's a bit more subtle · it vaguely resembles me (skull shape OK), but it turned into a stock photo. A frozen pose, a stock beige suit, a catalog smile, real-estate-ad lighting. If I put that as my LinkedIn photo, my network would wonder whether I pivoted into a Big4 consultant.

Why it happens · image models have a default voice. When you don't give them very precise instructions, they lean toward what they've seen most during training · editorial magazine photos, corporate photos, Hawkesworth-Kinfolk stuff for the outdoors. And it's never what you want when you're going for natural.

The fix that worked · I created a "natural iPhone" mode by default. The prompt forces things like "24mm lens, mixed ambient light, slightly off-center, no film grain, no editorial composition." Result · the photos finally look like what you'd shoot with a modern iPhone, not like a magazine cover.

Flop 3 · The editorial street shot that reeks of a Vogue shoot

At first glance, it's beautiful. Golden light, cream coat, yellow taxi in the background, a cinematic gesture of the phone to the ear · you'd think it cost €300 from a photographer.

Why I throw it out anyway · it's exactly the trap of flop 2, premium edition. Too editorial for personal use. If you put that on LinkedIn, your network immediately senses the "magazine shoot" and the photo loses all credibility of authenticity. It's pretty but it no longer looks like you · it looks like an Hermès ad from the 2023 season. The scripted movement (walking + phone) adds a layer of staging that reeks of a set.

The lesson · image models love magazine compositions. The "prettier" it is, the more suspicious you should be. A photo that really looks like you often seems less remarkable than a Vogue photo · and that's precisely why it works.

The 3 things I wish I'd known at the start

If you're thinking of building your own tool (or even just using a Klayn and friends), these 3 lessons will save you two weeks.

1 · Image models have a very magazine "default voice"

It's the most treacherous trap. You write a simple, neutral prompt and the model hands you Vogue. Not because you asked for it, but because it's what it saw most during training.

In practice · if you want natural iPhone, you have to impose it hard in the prompt (and even explicitly ban words like "Kodak Portra," "cinematic," "editorial"). If you don't do that, you'll always get fake-magazine and you won't understand why.

2 · Shorter = better (compress your prompts ×3)

My reflex at the start · add context. More words = more precision, right? No, the opposite.

Image models have an optimal word count between 30 and 100. Above that, they average things out and forget half your instructions. I compressed my prompts from 470 words to 165 words · the results were immediately better.

3 · No universal winning model · you test all 4 on your own face

There are about a dozen image models accessible through fal.ai. I mainly use 4 · Seedream 4.5 (the best for producing 5 consistent photos at once), Nano Banana Pro (the best for a single, very true-to-life photo), FLUX 2 (good for very precise outdoor scenes), GPT Image 2 (good for identity fidelity when you have a single reference photo).

None is better than the others in absolute terms · it depends on your face and what you're after. The only way to know is to generate the same scene with all 4 models on your selfies, and look. Budget $0.50 for that test, it's worth it.

Want to build it yourself?

If you're a developer or you have Claude Code installed, you can totally do this yourself. Here's the minimal stack:

• Next.js + Tailwind for the interface (3 pages, that's it · upload, scene choice, results)
• Zustand to store sessions in the browser (no database needed)
• fal.ai as the gateway to the image models (Seedream, FLUX, etc.) · pay-as-you-go billing, no subscription
• OpenRouter to call Gemini 2.5 Pro, which breaks the scene down into 5 shots · about €1 for 50 sessions

3 non-negotiable tips if you get into it:

1. Put a cost counter in from day 1. You're going to make mistakes, so you might as well see them right away. 30 lines of code, it costs you nothing.
2. Compress your prompts ×3. Start short, add only if you see it improving things.
3. "Natural iPhone" mode by default. Otherwise all your results will lean toward magazine and you won't understand why.

I wrote the full technical story (17 sections, every bug solved, every architecture decision) in a separate document. If you want to read it before you get started, tell me in a reply to the newsletter · I send it privately to anyone who asks.

Just want to use it? €39 · one-time purchase

You've got a team to kit out, or you want to redo your pro photos every three months without signing up for a €200/month subscription. I sell the source code · one-time purchase, no SaaS that can disappear, no per-user quota.

The 2-year math, no bullshit

	Klayn (SaaS sub)	Photographer (1/quarter)	My tool
Year 1	€2,400	€2,000	~ €130
Year 2	€2,400	€2,000	~ €30
2-year total	€4,800	€4,000	~ €160

A factor of 25-30, not a 10% saving. And you can redo 50 sets in a year at no extra cost — a new teammate joins? €1 of fal.ai credit and it's done.

Why not Klayn? Why not a photographer?

Three things that would have kept me up at night if I'd signed with Klayn:

• Vendor lock-in on your own face. Klayn trains a proprietary model on your selfies (4 h of training). You change your haircut, you shave your beard? You go back through the training step. With mine, 4 new selfies at each session, the AI adapts on the fly — 30 seconds, that's it.
• Dependence on the SaaS surviving. If they shut down, pivot or raise prices, your data leaves with them. With the source code, you have the files for life — even if I disappear tomorrow (unlikely, I sign with my name).
• The paywall on new models. Every 3-4 months a new image model comes out (Seedream 5, FLUX 3…). At Klayn = a new €299/month pricing tier to access it. With mine = one free commit that adds it to the dropdown.

And the photographer, then? It's complementary, not competing. A good session once a year is still valuable. But between two sessions, your set goes stale after 4 months — and for a team of 5-10, bringing a photographer to the office (€1,500-3,000 + a day blocked off) becomes irrational when everyone can send 4 selfies from their couch in 2 min.

Buy for €39 incl. tax →

Not a developer? The Settings page walks you through it step by step, you never touch a terminal. Write to me directly if you'd rather pay another way (bank transfer, etc.).

Going further

A few links if you want to dig into the topic:

1. fal.ai · the platform that lets you generate images through about a dozen models (Seedream, FLUX, Nano Banana, GPT Image, Imagen…). Pay-as-you-go pricing, topping up your account with €20 already gives you enough to test.
2. My Claude Code guide · if you want to get into building this kind of tool, that's where I usually start.
3. The AI glossary · to understand the technical terms (LoRA, polas, batch, queue) that come up often in the world of image models.
4. My tools · the 6 tools I use every day in my AI stack.

If you've built your own personal branding photo generator or tested Klayn, reply to the newsletter and tell me what you learned. That's how my next articles get better.

— Going further

3 articles to read next.

• Airbnb photos with Nano Banana — the other case where I replaced a pro shoot with AI generation, with the full prompt and cost.
• The sales tool built with Claude Code — another "put together solo without being a dev" project, useful for seeing the method applied to a different case.
• My Claude Code takeaways after 6 months — the workflow that makes this kind of little generator possible, without hand-coding.

— Questions people ask me

Frequently asked questions.