Extended metadata schema

The PhotoSpeak XMP-ps namespace — generated from the shared schema

PhotoSpeak stores everything it knows about a photo inside the photo file itself, as XMP metadata. The descriptive fields use standard namespaces (Dublin Core, IPTC, MWG regions); everything richer — voice recordings, transcripts, people, place, story, embeddings, provenance — lives in a custom namespace so it travels with the image and is yours to keep.

Namespace URI: http://photospeak.app/xmp/1.0/

Preferred prefix: ps (e.g. ps:Weather)

Free reader / recover your data: photospeak.app/viewer

Recovering your data without PhotoSpeak

Every enriched image also carries a plain-language manifest in the standard photoshop:Instructions field (visible in Photoshop, Lightroom, Bridge and Windows file details), naming this namespace and listing its fields. So even if PhotoSpeak is gone, the image tells you what it holds and where to read it. The data is standard XMP: any XMP library (for example the open-source ExifTool) can extract the raw ps:* values, then decode them as described below. The free Universal Gallery viewer does all of this in your browser, with no account and no server.

How the heavy fields are packed

Vectors (CLIP / face / voice embeddings) - little-endian float16 -> zlib deflate -> base64, stored as text. Each entry is tagged with its model id and dimension so it can be decoded without guessing.
Audio (transcription, context notes, voice threads) - Opus in an Ogg container, base64-encoded. Duration and a short SHA-256 hash travel alongside for integrity.
Images (photographer portrait, reversible watermark patch) - JPEG (portrait) or lossless PNG (watermark patch), base64-encoded, with a short SHA-256 integrity hash.
Structured data (regions, voice threads, mode data, ledgers) - UTF-8 JSON stored as text. Face/object regions also use the standard MWG RegionInfo structure so other tools can read them.

Every field in this namespace

This namespace defines 82 fields. Not every image contains every field — each carries only what applies to it.

Field	Type	Description
Context note (audio)
ps:ContextNote	string	Context note record identifier
ps:ContextNoteDuration	string	Duration of the context note audio
ps:ContextNoteFormat	string	Audio format of the context note source
ps:ContextNoteHash	string	Hash of the context note audio file
ps:ContextNoteText	string	Full text of the context note
Continuation sets
ps:ContinuationOf	string	pHash of the parent image this is a continuation of
ps:LinkedImages	string	JSON array of image UUIDs linked in this continuation set
ps:VolumeNumber	integer	Volume number in a continuation set (1 = original, 2+ = overflow)
Derivatives
ps:AnimatedFacePaths	string	Paths to animated face derivatives
ps:ColourisedPath	string	Path to the colourised version of this image
ps:DepthMapPath	string	Path to the depth map for this image
Detection
ps:DetectReviewedTypes	string	Detection types that have been reviewed
ps:DetectSuppressed	string	Detection results suppressed by the user
Diarization
ps:TranscriptionDiarization	string	Speaker diarization L1 block (JSON) for the transcription audio.
ps:VoiceEmbeddings	string	Per-segment packed voice embeddings for the transcription audio.
Embeddings
ps:FaceEmbeddings	string	Per-region face embeddings (model-tagged JSON array). See desktop xmp.py.
Enrichment control
ps:EnrichmentOptOut	string	When '1', the geocode/environment/weather pipeline steps skip this image so derived fields can be hand-curated.
Environment
ps:AirQuality	string	Air quality index at capture location
ps:DominantColors	string	Dominant colour palette extracted from the image
ps:Elevation	string	Elevation at capture location
ps:GoldenHour	string	Golden hour window at capture date and location
ps:MoonPhase	string	Moon phase at capture date
ps:NearbyPOIs	string	Points of interest near capture location
ps:Season	string	Season at capture date and location
ps:SunPosition	string	Sun altitude and azimuth at capture time
ps:Sunrise	string	Sunrise time at capture date and location
ps:Sunset	string	Sunset time at capture date and location
ps:Terrain	string	Terrain type at capture location
ps:Timezone	string	Timezone at capture location
ps:UVIndex	string	UV index at capture time and location
ps:Weather	string	Weather conditions at time and place of capture
Hashes
ps:ClipEmbedding	string	CLIP model embedding vector
ps:ClipEmbeddingHash	string	Hash of the CLIP embedding for integrity check
ps:ClipEmbeddingModel	string	CLIP model identifier used for embedding
ps:ImageDHash	string	Difference hash of the image
ps:ImageHashDimensions	string	Dimensions used when computing image hashes
ps:ImagePHash	string	Perceptual hash of the image
Identity / author
ps:AppVersion	string	PhotoSpeak app version that last wrote metadata
ps:PhotographerIdentityId	string	Links the author to a row in the identity library (unified people model).
ps:PhotographerPortrait	string	Embedded or referenced photographer portrait
ps:PhotographerPortraitHash	string	Hash of the photographer portrait data
ps:SourceNotes	string	Notes about the source or origin of the image (free-text journal entry)
Modes (knowledge substrate)
ps:Collections	string	RESERVED, DORMANT (do not reuse this tag). Was the planned per-image collection-membership codec (JSON array [{id, name}]). SUPERSEDED for curated Collections by the account-data model (Collections S2-S6, 2026-07-05..06): membership now lives in the DB + profile mirror keyed by ImageUUID, NOT in the image - see docs/adr/0007 (Collections half) + docs/current/Image_Grouping_And_Relationships_WIP.md. The tag name stays reserved as the likely foundation for Linked-images (Concept B) / share-in-file, so it is intentionally NOT deleted.
ps:ModeData	string	Knowledge substrate v2 (ADR 0007). JSON ARRAY of self-describing mode entries: [{mode, name, version, fields:[{k, label, type, unit?, value}]}]. Self-describing so any PhotoSpeak client or library reader interprets it WITHOUT the mode definition. Field option-lists/order/icon/lens-hints/capture-prompts live in the definition registry (profile/ mirror), NOT the image.
Provenance
ps:CopiedFrom	string	The parent instance's ImageUUID. Stamped when a same-machine duplicate (same UUID, different file) is renumbered to a fresh instance UUID, recording the lineage link (C2PA ingredient-reference pattern).
ps:ImageContentHash	string	Exact sha256 of the image's pixel bytes. Confirms a byte-identical copy (vs corruption) when a same-machine ImageUUID collision is detected. The exact UUID collision is the fork trigger; this is the copy-vs-corruption confirmation.
ps:ImageUUID	string	Unique identifier for this image instance
ps:OriginInstance	string	Instance identifier of the originating PhotoSpeak install
ps:OriginOwner	string	Owner identifier from the originating instance
ps:OriginPlatform	string	The platform on which this image's PhotoSpeak instance ORIGINATED (desktop/mobile/web). Set ONCE at mint time alongside ImageUUID + OriginInstance and preserved thereafter - a later edit on another platform does NOT change it (that is AppVersion / FieldHLCs). A fork mints a new instance, so it takes the forking platform; CopiedFrom preserves the lineage.
ps:ProvenanceChain	string	The HLC-ordered, hash-linked event ledger (JSON array of ledger entries) - the durable twin of the coordinator's sync_events. See docs/reference/Ledger_Entry_Schema.md. Supersedes the SyncVersion/PublishedVersion slots.
ps:PublishedVersion	integer	Version number when last published
ps:SyncVersion	integer	Monotonic sync version counter
Scene analysis
ps:EraMismatch	string	Flag indicating EXIF date conflicts with estimated era
ps:EstimatedEra	string	Estimated decade or era the photo depicts
ps:EstimatedEraScore	string	Confidence score for era estimation
ps:IsBlackAndWhite	string	Whether the image is black and white
ps:PhotoOfPhoto	string	Whether this image is a photo of another photo
ps:PhotoOfPhotoScore	string	Confidence score for photo-of-photo detection
Species
ps:Species	string	Species detected in the image
Subject (vs capture)
ps:SubjectDate	string	What the photo is OF (ISO-8601 partial e.g. 1960 / 1960-05), distinct from EXIF capture date.
ps:SubjectGPS	string	Lat/lon of the scene depicted, distinct from capture GPS.
Sync / merge
ps:FieldHLCs	string	{field_name: hlc} - the HLC of the last write to each scalar metadata field, so cross-device edits converge last-writer-wins on HLC instead of arrival order. Travels with the image as part of merge state.
Transcription (audio)
ps:Transcription	string	Audio transcription record identifier
ps:TranscriptionDuration	string	Duration of the transcription audio
ps:TranscriptionFormat	string	Audio format of the transcription source
ps:TranscriptionHash	string	Hash of the transcription audio file
ps:TranscriptionText	string	Full text of the audio transcription (human-editable; feeds the LLM extractor)
Transcription boundaries
ps:TranscriptionOffsets	string	Per-recording boundaries within the appended transcription audio.
Voice threads
ps:VoiceThread1	string	Voice thread slot 1
ps:VoiceThread2	string	Voice thread slot 2
ps:VoiceThread3	string	Voice thread slot 3
ps:VoiceThread4	string	Voice thread slot 4
ps:VoiceThread5	string	Voice thread slot 5
ps:VoiceThread6	string	Voice thread slot 6
ps:VoiceThread7	string	Voice thread slot 7
ps:VoiceThreadCount	integer	Number of voice threads attached to this image
ps:VoiceThreads	string	Voice threads container or summary
Watermark
ps:WatermarkPatch	string	Reversible watermark: lossless clean-crop patch that restores the marked region in PhotoSpeak/HTML viewers.
ps:WatermarkPatchHash	string	Integrity hash of the reversible-watermark clean-crop patch.
ps:WatermarkRect	string	Placement rectangle of the reversible watermark.
ps:WatermarkStyle	string	Style descriptor of the reversible watermark.