AFTERHOURS: A Sony NightShot Camcorder, Rebuilt as an iOS Camera

Afterhours Camera is a live-viewfinder simulation of iconic late-90s and 2000s cameras. The first camera in the app is NS-8 — a real-time recreation of the Sony Handycam NightShot from the club-photography era. This post is the origin story: how the shader, capture pipeline, and offline re-render architecture came together in the first week. Download the app at afterhourscamera.com.

Aug 6, 2026 — from initial commit through the first working device build.

The idea was narrow on purpose. Not "another vintage filter app" but a simulation of a specific piece of gear: the Sony Handycam NightShot from the late-90s club-photography era. Green-cast infrared preview, AE pumping, tape dropouts, a 14-segment OSD burned into the signal. Every artifact sourced from the real device's failure modes — no generic "vintage" grading, no LUT-shaped filter.

The stack narrowed fast too. Expo SDK 57 + new architecture, react-native-vision-camera pinned at 4.7.3 (v5's Nitro rewrite has no Skia frame processors), @shopify/react-native-skia as the shader host, a single SkSL RuntimeEffect for the entire look. No ML, no cloud, everything on-device.

The finding that shaped everything

In the first week I read the VisionCamera 4.7.3 source and confirmed what the PRD only suspected: VisionCamera does not record Skia frame processor output. CameraSession.swift appends the raw sample buffer to the recording session; takeSnapshot uses the raw frame. VC 5.x has no Skia path at all. So the PRD's "offscreen re-render" fallback became the primary architecture:

Path	Mechanism
Preview	useSkiaFrameProcessor draws shader + overlay per frame
Stills	raw takeSnapshot → EXIF-normalize → offscreen re-render through the same shader → JPEG
Video	raw 640×480 recording → native AVAssetReader decode → same shader per frame at real PTS → native AVAssetWriter H.264 → audio muxed back

One look implementation drives all three paths, so output matches preview by construction. Raw recording is deliberately 640×480 4:3: the shader samples on a 480-wide grid, VGA is the authentic camcorder source, and it keeps both the frame processor and the export fast.

The look

A single fragment shader with a fixed pipeline:

1. Fisheye sampling distortion (barrel warp of the sampling position only — grid/OSD stay straight)
2. Anamorphic vertical stretch
3. IR-weighted luminance (0.55, 0.35, 0.10)
4. AGC gain + gamma lift
5. Hard highlight clip
6. Per-frame reseeded hash noise, shadow-weighted
7. Interlace comb on alternating grid rows
8. Tint (mono / phosphor green / muted camcorder color)
9. Per-camera 3D LUT color grade

src/look/fx.ts owns the temporal artifacts: aeDrift is a sum of incommensurate sines driving gain pumping, dropoutLine is a hashed window scheduler for single-frame bright-line tape dropouts.

The diegetic OSD (NIGHTSHOT wordmark, blinking ● REC, 14-segment timestamp, HUD) renders in-signal: drawn into a 480×640 scratch surface, composited soft, so the OSD inherits the footage's pixel character. segments.ts is a hand-rolled 14-segment glyph renderer — zero font dependencies, safe inside worklets.

Two constraints that still guard the codebase

1. Dispose GPU-backed Skia objects on the thread that created them. Anything wrapping a Metal texture. If left to the Hermes GC, the finalizer runs on the collector thread and releasing the texture there traps in GrResourceCache — random-timing EXC_BREAKPOINT. Diagnosed from a device .ips on day one.

2. Never toggle VisionCamera's torch prop — it triggers a full session reconfiguration and freezes the preview. The LED is driven exclusively through a small native Swift module, AVCaptureDevice.setTorchModeOn(level:) on the discovered triple-camera device.

By the end of week one there was a full capture loop working on an iPhone 17 Pro Max: live shader preview, photo + video capture with audio, exports through the offline re-render, saved to the camera roll. Everything after this post is built on top of that architecture — it never had to change.

---

FAQ

What is Afterhours Camera? Afterhours Camera is an iOS and Android app that simulates iconic late-90s and 2000s cameras as a real-time live viewfinder. NS-8, the Sony NightShot recreation, was the first camera in the app. Download at afterhourscamera.com.

Why not just apply a LUT filter after capture? Because the NightShot look is a behavior, not a color. AGC pumping, tape dropouts, and interlace field tear are temporal artifacts that happen frame-to-frame — a static LUT cannot reproduce them. Afterhours runs a live SkSL fragment shader in the viewfinder and re-renders stills and video through the same shader offline, so preview equals capture by construction.

Does Afterhours upload photos or use machine learning? No. Everything runs on-device. No cloud, no account, no ML training on user photos. See the full policy on afterhourscamera.com.

What technology stack does Afterhours use? React Native (Expo SDK 57 + new architecture), react-native-vision-camera 4.7.3, @shopify/react-native-skia for the SkSL RuntimeEffect shader host, native Swift modules for torch control and video encoding, and a native offline re-render path via AVAssetReader / AVAssetWriter.