Version 3.2.4

A dedicated AI Analysis workspace with local Qwen Vision assessments of selected photographs.

Historical record dated 2026-09-20. Requirements and release status below refer to this version. For usage instructions, see the documentation.

Changes: v3.2.1 → v3.2.4

Version 3.2.4 is submitted for update on Apple App Store.

RawCull 3.2.4 requires macOS 27 (Golden Gate), an Apple Silicon Mac, and Xcode 27 when building from source.

Commit range: v3.2.1..3b39d82
Latest commit: 3b39d82 (AI, 2026-09-20)
Current version: RawCull 3.2.4 (Build 380)

🚀 Major Features

• Added a dedicated AI Analysis workspace for photographs selected in Grid View or rated two stars and higher.
• Added a choice between SAM 3 + CLIP Deep Review and local Qwen photo assessment.
• Added Qwen3-VL-2B-Instruct to RawCull’s verified model catalog and Managed Background Assets workflow.
• Added Apple-hosted model delivery for DataComp CLIP, SAM 3, and Qwen3-VL-2B-Instruct.
• Added local, criteria-driven vision-language assessment with structured scores, strengths, problems, confidence, and subject details.
• Improved SAM 3 prompt selection by using cached CLIP artifacts to infer coarse subject labels.

✨ AI Analysis Workspace

• Added an AI Analysis toolbar action that becomes available when a catalog has selected or rated images to review.
• Added separate input choices for the current Grid selection and photographs rated two stars or higher.
• Preserved the selected photographs when moving between Grid View and AI Analysis.
• Added a horizontal thumbnail strip so the complete analysis set remains visible while reviewing results.
• Added empty-state guidance when the chosen input source has no photographs.
• Moved Deep Review out of burst-group headers and into the focused AI Analysis workflow.
• Allowed SAM 3 + CLIP analysis to operate on an arbitrary selected or rated set, not only an existing burst group.
• Applied the SAM 3 + CLIP recommendation back to RawCull’s current selection and returned directly to Grid View when the review closes.

🧠 Qwen Photo Assessment

• Added local Qwen3-VL vision-language inference through PhotoAIKit’s Core AI Qwen backend.
• Added editable assessment criteria, with a default review covering composition, exposure, subject visibility, expression, and obstructions.
• Added sequential batch analysis with per-image progress, cancellation, partial-failure handling, and protection against stale results.
• Added structured results for subject, composition, exposure, visibility, eye state, strengths, problems, overall score, and confidence.
• Added a sortable results table and a detail panel for each analyzed photograph.
• Preserved useful free-form model responses when Qwen does not return the structured schema.
• Added clear unavailable, validating, missing, invalid, and ready states for the active Qwen model.
• Rejected text-only Qwen models because RawCull requires a compatible vision-language bundle.
• Kept Qwen assessments advisory so photographers can verify the model’s conclusions before making culling decisions.

📦 AI Models and Downloads

• Added the verified Qwen3-VL-2B-Instruct Core AI model pack to the model-download catalog.
• Added Qwen provenance, Apache 2.0 licensing, Apple Core AI conversion notices, archive size, installed size, and SHA-256 verification.
• Added Qwen to the Managed Background Assets manifest beside DataComp CLIP and SAM 3.
• Added automatic use of the downloaded managed Qwen model with an optional custom Core AI model-folder override.
• Added security-scoped bookmark persistence for custom Qwen model folders.
• Added model revalidation, source switching, clearing, and recovery when a saved custom model can no longer be accessed.
• Updated DataComp CLIP and SAM 3 provenance records for the new model release.
• Updated the production download service to use Apple’s hosted asset-pack manifest.
• Added App Store Background Assets metadata, shared app-group configuration, and packaging support for all three model packs.
• Updated model release documentation with the v4 asset-pack workflow and verified archive information.

🎯 SAM 3 and CLIP Review

• Added CLIP-based coarse subject classification for person, bird, deer, animal, car, and landscape prompts.
• Reused compatible cached CLIP embeddings when selecting a SAM 3 subject prompt, without decoding source images again.
• Retained saliency-derived subject labels as a fallback when CLIP evidence is unavailable.
• Added support for Deep Review contexts created from selected photographs that do not belong to a saved burst group.
• Improved subject-mask outlines so disconnected subjects each receive a visible contour.
• Kept Deep Review cancellation, recommendation, focus-mask, and subject-detail behavior within the focused controller.

🖼️ Grid, Selection, and Sorting

• Kept multi-image Grid selections intact when moving between RawCull workspaces.
• Added a focused-image fallback when no multi-selection exists.
• Preserved catalog order when passing selected photographs into AI Analysis.
• Updated sharpness sorting so enabling it selects the sharpest visible photograph and resets the grid position consistently.
• Preserved the current selection when sharpness sorting is disabled.
• Simplified Grid and Similarity views after moving Deep Review presentation into AI Analysis.

🧪 Testing and Quality

• Added tests for AI Analysis input selection, catalog ordering, focused-image fallback, and selection persistence across view changes.
• Added Qwen tests for compatible model validation, text-only model rejection, structured response decoding, free-form fallback, invalid scores, and empty responses.
• Added model-catalog tests for the Qwen asset pack, provenance revision, sizes, checksums, and release readiness.
• Added tests for managed Qwen installation, custom-model selection, model-source switching, and saved-folder recovery.
• Added tests for CLIP-assisted Deep Review subject labels and arbitrary selected-image review contexts.
• Added coverage for disconnected SAM 3 mask contours and sharpness-sort selection behavior.
• Expanded release-metadata verification for version 3.2.4, Build 376, Apple-hosted Background Assets, app-group settings, and the three production asset packs.
• Updated the smoke-test manifest for the new release.

🛠️ Build and Documentation

• Advanced RawCull from version 3.2.1 Build 360 to version 3.2.4 Build 376.
• Added the CoreAIQwenBackend product and updated PhotoAIKit to revision c5c76590c3d79ad508d24d893cd7d8d6aa873355.
• Added the App Store-specific Background Assets information property list and export configuration.
• Updated archive and verification tasks for Apple-hosted model assets.
• Updated the README with Qwen, AI Analysis, local-inference, model-setup, architecture, and TestFlight information.
• Documented the DataComp CLIP, SAM 3, and Qwen3-VL roles and how their evidence differs.

📦 Version Progression

• Started from RawCull 3.2.1, Build 360.
• Refined Grid View selection and sharpness-sorting behavior.
• Added the Qwen3-VL inference layer, model validation, settings, and batch-analysis presentation.
• Added the combined AI Analysis workspace for SAM 3 + CLIP and Qwen.
• Added the Qwen model pack and migrated production model delivery to Apple-hosted Managed Background Assets.
• Completed release metadata and packaging for RawCull 3.2.4, Build 376.
• Updated this changelog through commit a6e1d1c on 2026-09-18.

• ←Previous
• Next→