Facetune runs nearly 70 latency-critical AI features and services on fal Serverless, with performance up to 39% faster than before the migration.

Industry:
Consumer photo editing
Use-case:
Latency-critical generative photo editing on fal Serverless
Results:
Nearly 70 AI features and services on fal Serverless, with measured features running 9% to 39% faster than on the previous setup.

Facetune is one of the world's most known apps, with millions of MAUs who use Facetune to feel like their best version. Its newest features are generative: skin retouching, object removal, lighting, hairstyles, and instruction-based edits. Each feature is a diffusion pipeline that must return a result within seconds - to support a quick iteration cycle and a positive user experience.

The challenge

A Facetune edit starts with a tap, and the user expects a result almost immediately. Every delay adds up across a session, and speed can't come at the cost of identity preservation, realism, or image quality.

On its previous inference setup, the same team that builds Facetune's models also ran the serving stack: Kubernetes clusters, per-feature GPU capacity, scaling, and monitoring. Given the strategic nature of these tasks, the Facetune team decided to find a serving partner that could take on the load while meeting a strict latency bar - this is how fal become a part of Facetune's daily operations.

Preparing the ground

Facetune spent four weeks of infrastructure work turning its existing workers into fal workers. Once that foundation was in place, the integration itself went quickly, and feature after feature moved to fal.

Running on fal

  1. One app per feature. Each feature runs as its own fal app, with its own queue and autoscaling.
  2. Capacity that follows the traffic. Dedicated reserved capacity covers Facetune's baseline, and the same apps burst onto fal's shared fleet at peak, so Facetune no longer provisions GPUs by hand for every feature.
  3. Monitoring inside Facetune's stack. Facetune's Grafana scrapes fal's Prometheus metrics endpoint, placing request latency, queue depth, and GPU, VRAM, and CPU utilization next to Facetune's own application metrics. Scoped API keys mean the monitoring integration never carries deploy permissions.
  4. Shared alerting. Alerts designed with Facetune's engineers watch for traffic drops, requests queuing for a GPU, and server-error spikes, and post to a shared channel.
  5. Stable through the move. Reliability held through the cutover and the tuning work that followed.

Diagnosing a slow start

Right after migrations, some features took a hit in performance. All rides can start off bumpy, but the dedication and collaboration from the fal Serverless team is what made them the right partner. fal's team was responsive and creative, and worked side-by-side with Facetune's team to support a quick mitigation and further optimize runtimes. Today, every feature measured runs faster on fal than it did on the previous setup, between 9% and 39%.

FeaturePrevious setupfal, after optimizationNow vs previous setup
Vanish (object removal)1.00 s0.91 s9% faster
Skin retouching0.78 s0.70 s10% faster
Hair color1.81 s1.30 s28% faster
Lighting2.88 s1.76 s39% faster

Working together

Throughout the process, fal stayed close to Facetune's team. It joined the weekly investigation, gave Facetune control over traffic, and built the monitoring Facetune asked for.

"When latency and uptime matter, fal shows up. They tuned inference across datacenters, built the monitoring we asked for, and stayed close to our team every step of the way. They understand our requirements and their importance to our business, and operate as true partners and not only providers."
Gilad Schreiber, Engineering Manager, AI Features & Backend, Facetune

Facetune builds the models and pipelines and sets the quality bar. fal runs the serving layer underneath. The two teams keep working from the same data toward the same latency goal.

Customer Highlights
Dream Labs
went from first contact to a public robotics API on fal Serverless in 14 days
Dream Labs runs its existing container on reserved GPUs across the US, so every robot connects to a nearby GPU. fal’s World Model Accelerator cut round trip time by up to 35% on the same model and GPU.
Bria needed to bring new models to market faster while expanding developer adoption
Bria partnered with fal in 2024 to make its models accessible to developers through fal's inference infrastructure and model marketplace. Today, Bria's engineers can take a model from research-ready to publicly available in days rather than weeks, without managing the underlying GPU infrastructure.
fal serves 70 of Facetune's AI features and services
Facetune moved its generative features, from skin retouching to object removal, onto fal Serverless. Each feature runs as its own app, with reserved capacity for the baseline and burst capacity at peak. Every feature measured now runs up to 39% faster than on its previous setup.
AutoHDR
edits 25% of US real estate listings on fal Serverless.
AutoHDR’s editing models run on fal Serverless, where capacity follows the morning rush of shoots. Photographers get edited listing photos back in about 15 minutes, and a small team ships new versions as deployments, not infrastructure projects.

Ready to build with generative media?

Select option
Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.