Live-feed ML noise reduction (VideoToolbox)

[erik-sutton95]

Add optional live-feed noise reduction using Apple's VideoToolbox / ML path, as a viewing aid on the 720p SoftAP tap.

This is not camera-side NR, not wind NR, and not the existing “AI” upscaler.

Related: roadmap index

What already exists (easy to confuse)

Feature	What it is	Where
HEVC/AVC decode	VideoToolbox decompression	HevcDecoder.swift
Feed upscale Fast	Lanczos	Operator Setup → Link → Processing
Feed upscale Quality	MetalFX spatial	A13+
Feed upscale “AI”	VTLowLatencySuperResolutionScaler	iOS 26+, FeedSuperResolution.swift
Wind NR	Audio DSP @2	Camera write, unrelated

The AI path infers spatial detail. Help copy already warns it is a viewing aid, not evidence. There is no temporal/spatial denoise of the tap.

Workstreams

1. Naming (cheap)

• Stop calling super-resolution “AI” in Operator Setup so it is not mistaken for NR

2. Prototype VTB ML NR

• Inventory current VideoToolbox frame-processor denoise APIs on the deployment OS (iOS 17 baseline vs iOS 18 / 26 processors)
• Sit NR after decode, before LUT / scopes, with a kill switch
• Measure glass-to-glass delay, thermal, and whether WAVE/zebra still see the pre-NR tap (scopes must stay honest — NR is a look, not a meter source)
• Default off. Critical focus / false color should not depend on a hallucinated clean plate

3. Android

• Non-goal until iOS quality is proven. Android would need a different ML/GPU path.

Constraints

• Do not denoise before depacketize / decode
• Do not bake NR into recorded files
• Do not require a model download for the first on-device path if Apple ships an on-OS processor
• Keep the core free of VideoToolbox (Sources/OpenPocketViewCore stays portable)

Done when

Operator Setup has an explicit Feed NR toggle, scopes still read the raw tap, and a hardware note says whether it is worth leaving on for 720p Pocket HEVC.

[erik-sutton95]

Tracked as maintainer task #85.