AI Portrait: Launching Paid AI Photography at Global Scale
Context
Following the success of the AI Filter MiniApp, the next commercial question was whether users would pay meaningfully more for a higher-stakes AI product, not a fun stylisation effect, but a functional replacement for something they'd normally spend money on in the real world.
AI Portrait was that bet. The product used generative AI to produce professional headshots and artistic portrait photos: business profile pictures, LinkedIn-ready shots, and K-beauty-style portrait series. The value proposition was direct: get photos that look like they came from a professional studio, without spending significantly more in the real world on a professional studio shoot.
The product launched in two regional variants. The Western version (Professional headshot) targeted business-context headshots. The Asian version (Portrait) leaned into the K-beauty and Japanese portrait photography aesthetic. Both required users to upload 10–15 reference photos and select style packages, with pricing around 30 credits per style set.
One additional layer of complexity: I was simultaneously operating the same H5 MiniApp across two distinct apps. BeautyPlus (dominant in Asia and Southeast Asia, with a large free user base) and AirBrush (Pixocial's sister app with a primarily North American audience and a higher baseline willingness to pay). The two apps shared the same underlying product but had meaningfully different user demographics, spending behaviour, and funnel dynamics, which required managing promotions, effect prioritisation, and conversion strategy separately for each.
The Challenge
- The technology wasn't there yet, and we knew it. In 2023, generative AI portrait quality was improving fast but hadn't crossed the threshold where results were consistently convincing. A stylised AI filter that looked slightly off was forgivable. A "professional headshot" that distorted facial proportions, changed skin tone, or made someone look 20 years older was not.
- Getting the aesthetics right across completely different markets. A Western business headshot and a Japanese portrait series require different everything: lighting aesthetic, colour treatment, pose conventions, styling cues, even the concept of what "professional" means visually.
- Pricing was structurally high. At 30 credits per style set, AI Portrait was the most expensive MiniApp product we'd launched. The forgiveness threshold for quality issues was lower, and conversion required stronger intent signals than the AI Filter bundle model.
- Transitioning from 1.0 to 2.0 without losing revenue. AI Portrait 1.0 operated on a pure credits model. When 2.0 relaunched as a subscription benefit, it meant rebuilding the revenue model around subscription conversion rather than direct credit sales.
Pre-launch
Effect calibration as GTM work
Before any promotion, we spent significant time working with the designer team to calibrate the style outputs by region. This wasn't just aesthetic preference; it was a commercial prerequisite. The product couldn't launch until the results were good enough that users who paid wouldn't immediately feel deceived.
For the Western market, the benchmark was a credible LinkedIn headshot: neutral background, professional attire simulation, natural skin tone, proportional facial rendering. User feedback from early testing flagged consistent issues: necks rendered too long, facial features distorted at certain angles, AI-generated outputs that looked processed rather than photographed. Each became a calibration request to the designer team before we widened distribution.
For Japan and Korea, the aesthetic target was different: softer lighting, K-beauty colour grading, style options that mapped to the photo booth and portrait studio conventions those markets knew. We built separate style libraries for the Asian variant, including "macaron," "preppy," "blacknwhite," "sporty," and "y2k," and tested save rates and purchase rates by style to understand which directions the audience actually responded to.
The localization work meant two separate product surfaces with different effect ordering, different visual hierarchies, and different promotional creative. The conversion data justified the investment: the Japanese and Korean variant showed higher purchase rates per user than the Western version, despite lower total user volume.
Phase 1: AI Portrait 1.0
Earning trust in an unproven category
AI Portrait 1.0 launched in late September 2023. The user flow was more demanding than AI Filter. Users had to select gender, read photo guidelines, upload 10–15 photos, select styles, and pay before seeing any results. Every additional step was a potential exit point.
- Of users who entered the experience, only around 19% in Western markets and 13.5% in Japan and Korea clicked through to try the product, compared to 30% or more for AI Filter. The gap reflected a market still in its early stages. The consumer mindset around AI-generated portraits and professional headshots had not yet matured, and the pool of users genuinely willing to trust AI with something this personal remained small.
- The upload step was the steepest drop-off: around 50% of users who reached the photo selection page exited without completing the upload. Getting users to commit to an upload-heavy, paid flow proved significantly harder than expected. Most users either didn't have 10 to 15 suitable photos on hand, or found the quality guidelines too strict.
- Of users who completed the style selection page, only around 0.5% in Western markets and 0.6% in Japan and Korea completed a purchase, a low rate, but not surprising given how much the product asks of users before they see any result.
In the first weeks after launch, the Japanese and Korean variant generated higher ARPU per paying user than the Western version. The reason was behavioral: the Asian style library offered a wider range of distinct aesthetics, such as macaron, preppy, y2k, and blacknwhite, and users who were satisfied with one set tended to come back and purchase additional styles. Western users were primarily buying business headshots where the style variation was limited, and most purchases stayed within a single set.
Transition: AI Portrait 2.0
Restructuring around subscription value
AI Portrait 2.0 launched in late April 2024 with a fundamental model change: subscription users could generate AI portraits for free, removing the credits barrier for the majority of paying users.
The business logic was clear. Making AI Portrait a subscriber benefit increased the perceived value of the subscription, which we expected to drive new conversions from free users who wanted access. Before committing to this model shift, we modeled the expected server cost impact based on peak usage data from Portrait 1.0. The projection showed that even at higher generation volumes among subscribers, the infrastructure cost remained within an acceptable range, making it commercially viable to offer AI Portrait as a subscriber benefit without eroding margin. The short-term cost was that existing subscribers no longer needed to spend credits on portraits, but the long-term gain was strengthening the subscription flywheel.
Removing the credits barrier had an immediate effect on engagement: subscribed users' average generation count rose around 25% after the 2.0 launch. The product was being used more once the friction of paying per generation was gone.
Free-to-subscriber conversion remained the harder problem. Around 82% of AI Portrait users were free users paying with credits, while the remaining 18% were existing subscribers accessing the product as a benefit. Of the free users, only 2 to 3% on BeautyPlus and 7 to 8% on AirBrush chose to subscribe to unlock the benefit and generate for free, compared to 22% on AI Filter by Phase 3. Looking back at the 2.0 data, the core issue was not how the product was positioned. It was that the output quality was not yet good enough, which also meant that AI Portrait as a subscriber benefit carried limited persuasive power. Users who were not already convinced by the results were unlikely to subscribe just to access more of them, and the fact that most free users tried only one style set and did not return to purchase again supported that conclusion. AI-generated portraits in 2023 and 2024 could not consistently match what a real photography studio produced, and users could tell the difference. Until the technology improves to the point where the results are genuinely convincing, a meaningful portion of users will not pay for it regardless of how the product is packaged.
The split between AirBrush and BeautyPlus also surfaced a useful diagnostic. AirBrush's North American users had higher purchase intent at every stage of the funnel, not because of anything different in the product, but because of who they were and what they came looking for. BeautyPlus's larger user base made it better suited for testing effect appeal and broad promotional reach. Reading the two apps as distinct user populations rather than one combined number prevented us from misreading BeautyPlus's lower conversion rate as a product failure when it was primarily an audience composition story.
What I Learned
Quality is a go/no-go gate for functional AI products, not a nice-to-have.
For creative stylisation effects, an imperfect result is part of the experience. For a product positioned as a professional photography replacement, users apply a completely different standard. Pre-launch calibration work with the designer team wasn't a phase of the project; it was a prerequisite for the project making commercial sense at all.
Localization at the aesthetic level requires treating markets as genuinely different.
The Western and Asian portrait variants needed different style libraries, different promotional creative, different UI hierarchies. The fact that Japanese and Korean users showed higher purchase rates per session, despite the higher cultural sensitivity to portrait quality, was a direct result of having built something that matched their specific aesthetic reference points.
Restructuring a monetisation model mid-product is manageable when the long-term commercial logic is clear.
The 2.0 transition from credits to subscription benefit reduced short-term credit revenue predictably. The metrics that mattered were whether subscribers used the product more and whether it pulled free users toward subscription, both of which moved in the right direction.
The same product can behave like two different products depending on whose hands it's in.
AirBrush's North American users had higher purchase intent at every funnel stage, not because of anything we did differently in the product, but because of who they were and what they came looking for. Reading that split clearly prevented us from misreading BeautyPlus's lower conversion rates as a product failure rather than an audience composition story.
All metrics and performance data referenced in this case study reflect the 2023–2024 period. Sensitive figures have been anonymised or aggregated, and no proprietary or commercially confidential information is disclosed.