Skip to main content
Back

VisionCamera V5 Rewrites React Native's Camera Stack With 60x Faster Modules

A ground-up rewrite on Nitro Modules brings in-memory photos, RAW capture, multi-cam sessions, depth streaming, and a GPU-accelerated resizer.

Repo, Full Access
  • VisionCamera V5 ships as a full rewrite on Nitro Modules, dropping ~3,000 lines of JSI/C++.
  • New Constraints API replaces the Formats system, ending guesswork about which feature combos work.
  • Photos now return as in-memory objects with EXIF, no forced temp-file round-trip.
  • Adds RAW capture (DNG/ProRAW), depth streaming, manual AE/AF/AWB, Multi-Cam sessions, and 8K photos on iOS.
  • New GPU resizer in Metal/Vulkan hits ~5x speedup for ML preprocessing pipelines.
  • Modular packages: barcode scanner, location, resizer, Skia, and worklets each split out; custom native outputs no longer require forks.

The React Native camera stack just got a foundational overhaul. VisionCamera V5, the community's most popular high-performance camera library at 9.4k stars, has shipped a full rewrite that touches almost every layer of the pipeline. The headline change is a migration to Nitro Modules, but the more interesting story is what that migration unlocked: workflows that were previously either awkward or impossible in React Native.

According to creator Marc Rousavy, V5 is a new foundation for the library: rewritten to Nitro Modules, redesigned around a new Constraints API, and flexible enough to unlock use cases V4 could not accommodate. If you build camera features into a mobile app, this is likely the most consequential React Native release of the cycle.

Killing the hand-written JSI layer

V4 leaned on years of hand-written JSI and C++ glue code to bridge JavaScript with the native camera APIs. That code was fast, but it was also a source of race conditions, memory faults, and the occasional SIGSEGV. V5 rips out that entire layer and hands the job to Nitro, making the core runtime faster and type-safe end to end.

The Frame Processor implementation, once the most complex part of the codebase, is now written in plain Swift and Kotlin. That change alone cut around 3,000 lines from the project. Nitro also reshapes the API itself: a CameraDevice is now a HybridObject whose properties are fetched lazily from native on access rather than serialized upfront. Rousavy claims this is roughly 15x faster than Turbo Modules and 60x faster than Expo Modules, which translates into lower startup latency for camera-heavy screens.

The Constraints API replaces guesswork

The biggest usability change is the new Constraints API, which retires the old Formats system entirely. Anyone who has shipped a camera feature across iOS and Android knows the problem: iOS gives you a clean list of supported format combinations, while Android's Camera2 API exposes capabilities individually with no guarantee that combinations like 4K + HDR + 60fps will actually work together on a given device. V4 tried to paper over this by stitching together a fake formats list, and the result was often a camera session that silently downgraded resolution or crashed on startup.