RawCull

Current project version: 3.3.3.

Cull faster. Keep the sharpest frame. Keep your photos private.

RawCull is a native macOS app for reviewing and culling RAW photographs. It combines fast embedded previews, camera focus points, sharpness scoring, burst comparison, natural-language search, and optional on-device AI—without sending your photos to an external inference service.

Built for Apple Silicon with Swift 6 and SwiftUI.

Highlights

	Feature	What it gives you
⚡️	Fast RAW review	Scan catalogs and browse cached thumbnails or full-size embedded previews.
🎯	Focus intelligence	See Sony and Nikon AF points, focus masks, saliency, and sharpness evidence.
🏆	Burst ranking	Group similar frames and surface the strongest candidate with confidence and cautions.
🔎	Search by meaning	Find photos with natural-language queries using local CLIP embeddings.
🧠	Deep Review	Use SAM 3 to isolate the subject and compare detail where it matters.
🖼️	Combined Review	Review selected or tagged photos with retained images, findings, supporting crops, coverage limits, and a structured technical record.
✨	AI photo critique	Ask Qwen3-VL to assess composition, exposure, visibility, expression, strengths, and problems.
🔢	Objects analysis	Ask Qwen to suggest visible concepts, segment their instances with SAM 3, and assess a numbered review board.
⭐️	A complete culling flow	Tag, reject, rate, filter, compare, and persist your decisions.
📦	Flexible export	Export JPEG previews or copy selected RAW files with live rsync progress.

RawCull also reads core EXIF data, supports developed RAW previews, caches compatible analysis across sessions, and monitors cache usage and memory pressure.

Local AI

RawCull's AI features run locally on the Mac. Model downloads and validation are managed by the app; photographs, masks, prompts, and results stay on the device.

Model	Purpose	Used for
DataComp CLIP	Understands image/text similarity	Semantic search, visual similarity, burst grouping, and coarse subject labels
SAM 3	Finds where a prompted subject is	Subject masks, AF-point checks, and detail-aware Deep Review
Qwen3-VL	Describes and evaluates a photograph	Structured photo assessment against editable criteria

In AI Analysis › Objects, RawCull uses Qwen to identify photographically relevant concepts, SAM 3 to segment matching visible instances, and Qwen to assess those numbered subjects locally on your Mac. Specific concepts can be entered manually. Results describe matches for the requested concepts; they are not a guaranteed inventory of every object in the scene.

CLIP image embeddings are computed once and reused for later searches. SAM 3 masks and compatible analysis artifacts can also be cached. Qwen assessments are advisory: RawCull validates their structure, but the photographer remains the final judge.

Combined Review

Open Combined Review from the dedicated toolbar button beside AI Analysis, then choose selected or tagged images as the input.

The default Photo-first workspace pairs the retained overview with concise findings, coverage limits, and optional reviewed-region overlays. Inspect evidence opens the supporting crop and observations in the Evidence workspace. You can return to Photo-first while keeping the selected image and region.

Saved reviews remain available outside the current input selection. Review settings, retention cleanup, selection comparison, and run diagnostics are available through disclosure controls. The complete structured record can be inspected and copied as JSON.

Findings retain their evidence limits and conflicts. If synthesis fails, available observations are labeled as observations only. Missing retained images do not erase the report or measurements.

Workflow

Open a RAW catalog
       ↓
Review previews, metadata, focus points, and sharpness
       ↓
Group bursts or search the catalog by description
       ↓
Compare candidates with Deep Review, Qwen, or Combined Review
       ↓
Rate, tag, reject, and export the keepers

Requirements

• Apple Silicon Mac
• macOS 27
• Xcode 27 and Swift 6 for development

RawCull is focused on Sony ARW workflows. Its parsing layer also contains Nikon MakerNote support for normalized AF-point extraction.

Architecture

The app keeps UI, workflow, caching, persistence, and culling policy in RawCull while focused Swift packages own reusable functionality:

Package	Responsibility
RawParserKit	RAW discovery, metadata, embedded JPEGs, previews, and MakerNotes
PhotoAnalysisKit	Sharpness, focus masks, saliency, classification, and calibration
PhotoAIKit	CLIP, SAM 3, Qwen, model validation, similarity, and mask storage
RawCullCore	Shared catalog, EXIF, burst grouping, ranking, and review models
RsyncArguments + RsyncProcessStreaming	Safe copy configuration and streaming execution
DecodeEncodeGeneric	Codable persistence helpers

Remote dependencies are pinned in Package.resolved. The current resolved versions are:

Package identity	Resolved pin
coreai-models	1.0.0
decodeencodegeneric	1.0.0
eventsource	1.5.1
parsersyncoutput	1.0.0
photoaikit	1.0.0
photoanalysiskit	1.3.1
rawcullcore	1.1.2
rawparserkit	1.3.1
rsyncarguments	1.0.0
rsyncprocessstreaming	1.0.0
swift-asn1	1.7.3
swift-collections	1.7.3
swift-crypto	4.5.2
swift-huggingface	1.0.0
swift-jinja	2.5.1
swift-transformers	1.3.5
xgrammar	0.2.2
yyjson	0.12.0

Model manifests, licence notices, and provenance records live in ModelAssets.

Build

Build and export a Debug archive:

make debug

Or build the Xcode scheme directly:

xcodebuild \
  -project RawCull.xcodeproj \
  -scheme RawCull \
  -destination 'platform=macOS,arch=arm64' \
  -onlyUsePackageVersionsFromResolvedFile

Test

make test-smoke        # Fast integration and critical-path coverage
make test-full         # Full suite with Thread Sanitizer
make test-performance  # Performance and extreme-concurrency coverage

The Swift Testing suites cover RAW parsing, focus and sharpness metrics, similarity and semantic search, Deep Review, Qwen responses, downloads, caches, persistence, cancellation, concurrency, and copy workflows.

Release

Validate the AI boundary and release inputs before building:

make verify-ai-import-boundary
make release-preflight
make build

make build creates the signed, notarized, and stapled app and DMG. It requires the configured Developer ID identity, notarytool keychain profile, and create-dmg at ../create-dmg/create-dmg.

After publishing, verify the downloaded artifact against the generated SHA-256 file:

make verify-downloaded-dmg \
  DOWNLOADED_DMG=/path/to/downloaded/RawCull.3.3.3.dmg

App Store Connect and TestFlight uploads

Build, sign, and upload in one command:

./Scripts/release.sh internal   # Internal TestFlight testing only
./Scripts/release.sh appstore   # TestFlight and eligible for App Store submission

The Makefile equivalents are make upload-internal and make upload-app-store. Sign in to your Apple Developer account in Xcode Settings > Accounts first. Both commands use the AppStore configuration, automatic signing, and pinned package versions. They preserve archives and export diagnostics under build/releases/. They do not submit for App Review or publish the app.

Preview either command by adding --dry-run. Xcode manages upload build numbers by default. To specify one explicitly for both the app and its extension, use BUILD_NUMBER=400 ./Scripts/release.sh appstore; choose an unused build number. For API authentication, set ASC_KEY_PATH to your .p8 file and also set ASC_KEY_ID and ASC_ISSUER_ID. Keep the key outside the repository.

After Apple processes the upload, check the build in App Store Connect and assign it to an internal TestFlight group if automatic distribution is not enabled. An internal upload cannot be used for external TestFlight or App Store submission; use appstore if you want to submit that same tested build later.

Local release integration tests

Run make releaseaiobjectstest to analyze the ARW files directly in Downloads using installed Qwen and SAM 3 models, without launching RawCull's UI. A run report identified by UUID is written as RawCull-AI-Objects-<UUID>.md in Downloads. See release integration instructions for model paths, overrides, and result semantics. The linked guide also contains historical baseline details; use the current Makefile target above for AI Objects.

Additional opt-in checks are available:

make releasesharpnesstest  # Catalog sharpness report; no Qwen or SAM 3 required
make combinedinputstest   # Combined Review input-contract probes with installed models
make combinedselectiontest # Six-file Sony source/crop probe and selection comparison

The catalog-based checks default to $HOME/Downloads; override it with RELEASE_CATALOG=/path/to/photos. AI Objects defaults to model bundles under $HOME/ModelAssets/Release/Models; override them with RELEASE_MODELS, RELEASE_QWEN, or RELEASE_SAM3. Input-contract probes use INPUT_PROBE_MODELS and write to INPUT_PROBE_OUTPUT (default /tmp/rawcull-combined-input-probe).

License

RawCull is available under the MIT License. The optional AI models retain their respective licences and attribution requirements; see ModelAssets/Notices.