PixelPatch scaled AI photo enhancement to 4M requests a day while cutting infra spend 40%

PixelPatch is a consumer photo app with a simple promise: tap once, and an old, blurry, or badly lit photo comes back restored. Behind the tap is a pipeline of diffusion-based enhancement models — denoising, face restoration, upscaling, and colorization — that turned the app into a word-of-mouth hit. Within eighteen months of launch, PixelPatch was processing 4 million enhancement requests a day from users in 140 countries.

The challenge

Consumer traffic gave PixelPatch two problems at once. The first was shape: daily peaks ran 5x overnight troughs, and viral spikes — a TikTok trend, a holiday weekend — could triple traffic in an hour. The team's hyperscaler deployment was provisioned for peaks, which meant paying peak prices around the clock. The second was unit economics: on premium-priced hyperscaler GPUs, the cost per enhancement was high enough that the free tier — the app's entire growth engine — lost money on every tap.

As the app grew, infrastructure became the largest expense in the company and the biggest threat to it. The founders gave the platform team a blunt target: cut GPU cost per request nearly in half without letting latency slip, because in a consumer app, a slow tap is an uninstall.

Why Reviosa

PixelPatch rebuilt its serving fleet on Reviosa L40S instances at $0.69/hr. The team's benchmark found the L40S hit the exact price-performance point their models needed: 48 GB of VRAM comfortably held the full enhancement pipeline with room for large inference batches, at a fraction of the hourly cost of the cards they were leaving behind.

The fleet design leaned into Reviosa's hourly billing. An autoscaler tracks the request queue and walks capacity up and down through the day — roughly 120 instances at the daily peak, under 30 in the overnight trough. The heaviest premium feature, a high-resolution restoration mode, routes to a smaller pool of on-demand H100 instances at $2.49/hr that scales independently. Nothing is reserved for a worst case that happens twice a year; when the spike comes, the autoscaler simply keeps walking.

"We load-tested the migration by replaying our single biggest viral day at 2x speed. The fleet scaled from 40 to 190 instances in under 25 minutes and median latency never left the 2-second band. That was the moment we stopped worrying." — Renata Vasquez, Head of Infrastructure

The results

Six months after the migration, with traffic up 30%:

• 4 million requests a day served on an elastic L40S fleet that tracks the traffic curve hour by hour.
• 40% lower infrastructure spend, despite the traffic growth — cost per enhancement fell by nearly half.
• 1.9-second median enhancement time, an improvement on the old fleet, with p95 holding steady through three viral spikes.

The unit-economics fix changed the business. The free tier now runs at a sustainable cost per tap, which let PixelPatch stop throttling free enhancements — a limit that had been suppressing the viral sharing loop the whole growth model depends on. Conversion to the premium tier rose 22% in the quarter after the limits came off, with premium users flowing to the H100 pool whose costs scale exactly with the revenue they generate.

PixelPatch's next act is already running on Reviosa: a video-restoration feature, currently in beta, trains on burst 8×H100 nodes and will serve from the same elastic pattern that carried photos to 4 million requests a day. This time, the team modeled the launch economics in an afternoon — because for the first time, they knew precisely what a request costs.