CASE STUDY · PHOTO AI

From a curiosity about AI to US$105,000 a month, in one PHP file

Pieter LevelsPieter Levels64 products · see the profile →10 parts · 19 sources · updated 13 Oct 2026 · not reviewed by Pieter

Written by MRRMaker from Pieter's own posts, his product pages and his interview on the Lex Fridman Podcast, read in full. Every fact links to its source. Not reviewed by Pieter.

01

At a glance

What it isA photo studio without a photographer: you upload photos, it trains a model of you, and you create photo shoots anywhere 1
Started asAvatar AI, a side project that sold US$100,000 in its first 10 days 2
RevenueUS$61,808 a month with 1,872 paying customers (July 2023) 3; US$105,000 a month, US$80,000 profit (March 2026) 4
TeamPieter alone, plus one AI developer for 10 months 5
StackRaw PHP and jQuery, no frameworks 3; models trained and run on Replicate 6
Price todayUS$19, 49, 99 or 199 a month; yearly plans give about six months free 7
02

How he found the idea

Pieter did not start from a problem. He started from curiosity: when Stable Diffusion came out, he installed it on his Mac and posted what it made on X. 8 It was bad at people but good at houses, so he made thishousedoesnotexist.org. 8

  1. Housesthishousedoesnotexist.org: the model was good at houses, bad at people 8
  2. Interior AIUS$10,000 to 20,000 a month within a week 9
  3. Avatar AIUS$100,000 in its first ten days, October 2022 2
  4. Photo AIlaunched on 10 February 2023 1
From houses to Photo AI, following what the model was good at.

People got bored of the houses, but the model was good at interiors. With a new technique called image-to-image he could upload a photo of a room and restyle it. That became Interior AI. Within a week it made US$10,000 to 20,000 a month, and two years later it still made US$40,000 to 50,000. 9

To improve Interior AI he learned fine-tuning: training an existing model on one goal. For fun he trained it on photos of himself, and it learned his face. "So I was like, this is another startup." That was Avatar AI. 10

His thesis at the time, posted in December 2022: the real opportunity in AI for most people is not the AI itself but a front end around it. 11

03

The first launch: US$100,000 in ten days

Avatar AI launched at the end of October 2022. Ten days later it had made US$100,000 from 2,943 sales, about US$33 each. 2

Then the margin vanished. The company that ran his model training saw the revenue he posted and raised its price from US$3 to US$20 per training. He was selling at about US$30, so the profit was gone. 12

Price of a saleabout US$30 12
Training, beforeUS$3 12
Training, afterUS$20 12
The training price rose from US$3 to US$20, against a sale of about US$30.

He went looking for another platform and asked Replicate's CEO to add the training method he needed. The answer was no, because it took too long on their GPUs. He kept asking, and a week later they launched it. By his account, the avatar apps that followed mostly ran on Replicate too. 6

04

The pivot: from a fad to a real problem

Pieter saw the avatars as hype. Cartoon and painted styles hide the AI's mistakes on a face; a real photo does not. He decided to solve the hard version: realistic photos, without a photographer or a studio. 13

It took two months, and he admits he was procrastinating because the project felt endless. Marc Kohlbrugge pushed him to ship. Photo AI launched on 10 February 2023. 1

05

Building it: scrappy on purpose

The first version had no real login. You paid through Stripe Checkout and received a login link. Proper Google login came a year later. His rule: decide what not to build, because there is no time. 14

One file, no framework.

By July 2023 the whole product was almost 14,000 lines of raw PHP with inline HTML, CSS and JavaScript, using jQuery and no other framework. It had 1,872 paying customers and made US$61,808 a month. 3 In March 2026 it was one file, index.php, of 40,870 lines. 4

He cannot write Python, which most AI work needs, so he hired one AI developer, who worked with him for 10 months. 5

06

Getting better without guessing

With more than a million photos generated a month, he stopped judging settings by eye. He tested different generation settings on 10% of users and measured which photos they marked as favourites or downloaded. The winning settings became the default. 15

The training data mattered more than training time. Face crops alone lost the body's shape, so he mixed full-body photos with face crops, automatically. 16

07

Growing the product

In June 2023 he added a try-on mode: upload clothes, and your AI model wears them. He aimed it at Shopify sellers who need model photos, at US$29 a month, which he called at least 30 times cheaper than a shoot. 17

In August 2026 he added a full video editor that generates scenes with the user's trained models. 18

Pricing in October 2026 has four monthly plans: Starter US$19, Pro US$49, Max US$99 (marked most popular) and Ultra US$199. Only Pro and above allow commercial use. 7

08

Running it: disputes can kill the account

If more than 1% of payments are disputed, Stripe, Visa and Mastercard can shut the account for good, and each dispute costs a US$30 fee, returned only if you win. For a decade he ignored disputes and lost almost all of them. 19

In 2026 he had AI help him build a responder. When a dispute arrives, the site collects what the customer did in the app and sends a detailed PDF as evidence. The first big win was a US$1,199 dispute from a customer who had generated thousands of photos. 19

09

Decisions you can copy

  • Build a front end around a model someone else trains. He never built AI models; he built products on top of them. 11
  • Ship before it feels finished. A login link instead of a login system, for a whole year. 14
  • Do not depend on one supplier's price. A price rise from US$3 to US$20 erased his margin overnight; he moved to a platform he had asked to add the feature. 12
  • Leave the fad for the real problem. Avatars were hype; photos without a photographer are a need. 13
  • Let users choose the best settings. Test on 10% of them and measure what they keep. 15
  • Sell the same engine to a new buyer. Try-on offered the same models to shop owners who need product photos. 17
  • Automate the boring defence. Dispute evidence built from usage data keeps the account under the 1% line. 19
10

What the record does not say

  1. How much does a user cost in GPU time, and how did the margin move from 2023 to 2026?
  2. Which plan brings most of the revenue, and how long does a subscriber stay?
  3. Where do new customers come from today: X, search, or paid creators?
  4. Why a monthly subscription, when Avatar AI sold one-time packs?
11

Sources

  1. Pieter, "My new project Photo AI", 10 Feb 2023 ↗
  2. Pieter, "Sold $100,000 in AI-generated avatars", 7 Nov 2022 ↗
  3. Pieter, "Photo AI is almost 14,000 lines of raw PHP", 3 Jul 2023 ↗
  4. Pieter, "Photoai.com is a 40,870 line index.php", 6 Mar 2026 ↗
  5. Lex Fridman Podcast #440: hiring an AI developer ↗
  6. Lex Fridman Podcast #440: asking Replicate for DreamBooth ↗
  7. Photo AI pricing page, read 12 Oct 2026 ↗
  8. Lex Fridman Podcast #440, 20 Aug 2024: the origin story ↗
  9. Lex Fridman Podcast #440: Interior AI ↗
  10. Lex Fridman Podcast #440: training on his own face ↗
  11. Pieter, "The real opportunity in AI", 10 Dec 2022 ↗
  12. Lex Fridman Podcast #440: the US$3 to US$20 price rise ↗
  13. Lex Fridman Podcast #440: why real photos ↗
  14. Lex Fridman Podcast #440: the login link ↗
  15. Lex Fridman Podcast #440: testing settings on users ↗
  16. Lex Fridman Podcast #440: training data ↗
  17. Pieter, "Photo AI adds try on mode", 14 Jun 2023 ↗
  18. Pieter, "I vibecoded a full video editor into Photo AI", 1 Aug 2026 ↗
  19. Pieter, "Winning my first big Stripe dispute", 30 Apr 2026 ↗