The Challenge

A finished listing photo is the edited image buyers see first on a real estate listing site: bright, evenly lit and ready to show the home at its best. A photographer shoots each room at several exposures, and AutoHDR blends them into one evenly lit image, then makes whatever edits the agent asked for. AutoHDR first ran its models on its own GPU cluster, so as the product grew, the team planned capacity for every new edit type and every busy morning of shoots itself.

The Solution

AutoHDR moved its editing inference off its own cluster and onto fal Serverless, where its models deploy as endpoints that fal scales with traffic. The team tests new functionality next to the ones in production, so a new version is an easy deployment, not a new infrastructure project.

“We aim to spend time strictly on the core competencies we’re uniquely positioned to do, and none of our time on the required work that doesn't widen our competitive moat. Partnering with Fal to scale inference offloads the overhead of doing it ourselves, so we can spend time on the engineering work that drives revenue.”

Matt Gia, Founder, AutoHDR

AutoHDR also calls fal’s hosted models from the same stack, so a small team covers a long menu of functionality without running a GPU fleet. At the volume of a quarter of US listings, that matters in a practical way: capacity follows the morning rush, and the team no longer plans it.

“This quarter, we've shipped 12 new models, trained on the portfolios of the 12 best photographers in the country. To move fast and be the most competitive company in our niche - this is the work we need to spend our undivided attention on. This quarter, FAL’s team and infrastructure has helped us do that.”

Matt Gia, Founder, AutoHDR