Wedding photo app ditches cloud APIs to run AI upscaling privately on own server

Knipsmig, a QR-code-based photo sharing service for weddings and parties, has added a local 4x image upscaling feature to improve low-resolution guest uploads before they appear in printed photo books. The developer chose to run the Real-ESRGAN model (realesr-general-x4v3) entirely on-premise rather than use third-party APIs like Gemini or OpenAI, citing three key reasons: generative models alter faces rather than faithfully adding pixels, privacy regulations require updating data processing agreements whenever a new vendor handles guest photos, and per-image API costs become significant at scale. The chosen model is a compact 5 MB ONNX file with around 1.2 million parameters, running roughly 10 times faster on CPU than heavier alternatives. It was integrated into the existing Ruby on Rails editor pipeline via the onnxruntime gem, processing images in tiled 512-pixel segments with overlapping borders to avoid visible seams. Key optimisations included bypassing Ruby array marshalling for tensor data, instead writing raw bytes directly into OrtValue objects, which dramatically reduced overhead beyond the inference step itself.