face_detection_tflite

• Default precision is now Precision.fp32 instead of fp16. This changes numeric output. flutter_litert 3.8.0 changed its own default for the same reason: across 29 published detection models measured on five GPUs, fp16 matched a plain-CPU reference for only about a fifth of them, while fp32 matched every model that compiled. These graphs emit pixel-space coordinates and landmark positions, and fp16 carries about three decimal digits of mantissa, so the error lands directly on output geometry. The cost is real and worth stating plainly: fp32 is a median 29.9% slower on GPU across those five GPUs, with Apple M4 the lone exception at 6.5% faster. Pass precision: Precision.fp16 explicitly to restore the previous behaviour, ideally per model and validated on your target GPU.
• Pin flutter_litert to ^3.8.0.
• Rename the example app's engine badge from CM / XNN to CM / Interpreter. XNNPACK is only the delegate the Interpreter path uses on desktop and Android; on iOS that path runs the Metal delegate, so an XNN label was wrong there. The badge switches between the two flutter_litert engine classes, CompiledModel and Interpreter, so it now names those. The button is fixed-width so swapping labels does not shift the surrounding controls. Example-only change; no library code is affected.
• Feature: opt-in temporal face tracking (FaceDetector.create(enableTracking: true)) assigns stable Face.trackingId values across sequential native and web detections. Motion-aware geometric association preserves IDs when detector ordering changes and across short detector dropouts; resetTracking() clears state when switching streams. maxMissedFrames (default kDefaultMaxMissedFrames, 3) sets how many processed frames a face may go undetected before its ID is retired, counted in frames the detector actually ran rather than wall-clock time, so a camera loop that drops frames while inference is busy can raise it; negative values throw ArgumentError before any model loads. Tracking-enabled calls are sequenced in invocation order, and combined detection + segmentation results are tracked too. Tracking is not face recognition; default behavior remains unchanged with null IDs.
• Export kDefaultMinFacePresenceConfidence (0.5) from both entry points. It has been referenced from the public dartdoc on FaceDetector.create() and initialize() since 6.7.0, but was never exported: the native entry listed only the model-name constants from face_model_config.dart, and the web entry did not export that file at all. The doc links therefore dangled on pub.dev and callers could not name the default they were being documented about. Additive only; no behaviour change, and the value is unchanged at 0.5. A test now guards the export so it cannot be dropped again.

• Face-presence gate (MediaPipe min_face_presence_confidence): FaceDetector.create() / initialize() now accept minFacePresenceConfidence, which drops detections the face-landmark model does not confirm as a face by gating the mesh "face flag" (face.meshScore). This is MediaPipe's standard second-stage check and suppresses common first-stage false positives such as a hand or palm, which clear the BlazeFace detector but score near zero on the mesh model. It defaults to 0.5, matching MediaPipe (unlike minScore/minFaceSize, which default to 0.0), so upgrading turns the check on: in standard/full modes, detections whose meshScore is below 0.5 are no longer returned. Pass minFacePresenceConfidence: 0.0 to restore the previous "return every detected box" behavior. The gate has no effect in fast mode (no mesh is computed), and a null meshScore always passes. On both native and web it runs right after the mesh stage, before iris and blendshape, so rejected faces skip that per-face landmark cost. Validated to [0.0, 1.0] (out-of-range or NaN throws ArgumentError). See the README "Detection Gates" section.
• Match MediaPipe's score_clipping_thresh exactly: the BlazeFace raw-logit clip limit (kRawScoreLimit) is now 100.0 (was 80.0), matching the upstream TensorsToDetectionsCalculator. This is numerically inert (sigmoid(80) and sigmoid(100) are both 1.0 in float32), so detector scores and which faces are returned are unchanged; the constant is aligned purely for exactness.
• Fix (web): activeAccelerator chained the model runners with ??, but every runner reports a non-null backend once initialized, so the chain always short-circuited on the detector model and ignored the other four. Runners compile independently and can fall back from WebGPU to WASM on their own, so when the detector is the one that falls back the aggregate reported wasm while other runners were still on the GPU. Both the runtime GPU-error fallback and the slow-WebGPU warmup are gated on that value, so neither would fire for the runners still on WebGPU. Now uses aggregateActiveAccelerator from flutter_litert, which reports webgpu if any runner is on it. A mixed state was observed live on Chrome (blendshapes on WASM, the rest on WebGPU).
• Add FaceDetector.acceleratorReport, a per-runner map of which backend each model actually compiled to, for diagnosing mixed WebGPU/WASM outcomes.
• Adopt the shared flutter_litert 3.6.0 helpers in place of local copies: compiledModelFromBufferAuto for the {gpu, cpu} accelerator branch, compiledFloatCount / compiledSquareInputSide for compiled tensor IO at 13 call sites, and collectOutputShapes for output shape collection. The local compiled-IO helpers returned a zero or negative element count for a degenerate tensor where the shared ones throw; a test asserts every bundled model reports positive float32-aligned tensor sizes, the domain where the two agree, so no model shipped here changes behaviour.
• Deprecate OutputTensorInfo, collectOutputTensorInfo and testCollectOutputTensorInfo. Both call sites only ever read the shapes and discarded the buffers; use collectOutputShapes from flutter_litert.
• Remove the web_image_utils re-export shim and import from flutter_litert directly.
• Update flutter_litert -> 3.6.0.
• Expand the README live camera section with the full production pipeline (frame throttling, orientation handling, cover-fit overlay mapping).

• Update flutter_litert -> 3.5.1.
• Performance (web): the minScore/minFaceSize gates now filter detections before the per-face mesh, iris and blendshape stages (as on native), and detectFacesWithSegmentation decodes the image once instead of twice. In two interleaved 60-run A/B pairs on Chrome 149 (WASM), using a warmed threshold that retained exactly one face from a 4-face group shot, full mode dropped from 65.8-66.8 ms to 46.1-46.6 ms per call (about 30% faster) and combined detection plus segmentation from 96.8-97.7 ms to 69.2-70.7 ms (about 28% faster); ungated detection was unchanged. Detector-level outputs (scores, boxes, keypoints, which faces are returned) are bit-identical; mesh-stage values for early invocations can shift within the web runtime's pre-existing call-order jitter, which is smaller than the jitter that runtime already shows between identical calls in unchanged code.
• Fix (web): a BlazeFace candidate whose decoded box was degenerate (nonpositive width or height) shifted every later candidate's box onto the wrong confidence score, corrupting weighted NMS and the reported face.score/minScore gating for those detections. Candidate decode now keeps each box paired with its own score (decodeBlazeFaceCandidates, pure Dart and unit-tested), and NaN scores remain rejected; results are unchanged whenever no candidate was skipped, which is the common case. Native was not affected. Because the fix can change web detection output for affected inputs, FaceDetector.modelVersion is now 1.1.1.
• Performance: face meshes returned by the native pipeline now keep their landmark data in the packed float buffer that crossed the isolate boundary and build Point objects lazily on first access (FaceMesh.packed). Callers that never read mesh.points (for example apps that only draw bounding boxes from full-mode results) skip 468 allocations per face per frame; callers that do read them get bit-identical values. Measured ~1.4% faster multi-face full-mode detection and ~3.5% faster adjacent embedding calls (less GC churn), pooled over 200 runs per side; single-face detection within noise; memory usage is equal or lower.
• Performance: embedding requests (getFaceEmbedding, getFaceEmbeddingFromMat, getFaceEmbeddingFromMatBytes, getFaceEmbeddings) now send only the two eye landmark points to the detection isolate instead of serializing the whole Face (468-point mesh and iris data included), and embedding vectors return as typed data instead of boxed lists. The eye points are exactly what face.landmarks reports (iris-refined when available), so embeddings are bit-identical. Measured ~4% faster per getFaceEmbedding call (3.41 ms to 3.28 ms median over 100 runs, Apple Silicon, XNNPACK).
• Performance: on native platforms, minScore/minFaceSize gates now run inside the detection isolate right after the detector stage, so gated-out faces skip the per-face mesh, iris and blendshape work instead of being computed and then discarded. Detection results are byte-identical to the previous late filtering; only the wasted per-face stages are skipped. In a benchmark on a 4-face group shot with a minFaceSize keeping one face (full mode, Apple Silicon, XNNPACK, median of 100 runs), latency dropped from ~18.0 ms to ~7.0 ms per call (about 61% faster). Ungated calls are unchanged. Adds the shared boxVisibleWidthFraction and applyDetectionGates helpers; Face.widthFraction now delegates to the former with bit-identical arithmetic.

• Update flutter_litert -> 3.5.0

• Update flutter_litert -> 3.4.1 (web CompiledModel WebGPU compile watchdog: a compile attempt that never settles now falls back to WASM instead of hanging). No API change.

• Update flutter_litert -> 3.3.1 (Android Gradle Plugin 9.x build fix; faster Interpreter.run/CompiledModel.run and fewer per-frame allocations in the camera YUV path). No API change.
• Head pose: Face.headEulerAngles (and headEulerAngleX/headEulerAngleY/headEulerAngleZ) report pitch, yaw and roll in degrees, following Google ML Kit's sign conventions. Pitch/yaw come from the 3D mesh (standard/full); fast mode gives roll only. Computed on demand, so no added inference cost.
• Face classification (MediaPipe Blendshape V2, full mode): Face.smilingProbability, leftEyeOpenProbability and rightEyeOpenProbability (ML Kit semantics, subject-relative left/right), plus blendshapes with all 52 coefficients indexed by the new Blendshape enum. Bundles face_blendshapes.tflite (955 KB, Apache 2.0); it is a CPU-pinned MLP validated against MediaPipe's golden output, with no cost in fast/standard (values are null there). See the README "Face Classification" section.
• Named face contours (Google ML Kit FaceContourType parity): Face.getContour(type) and Face.contours return ordered mesh points for the face oval, eyebrows, eyes, lips, nose and cheeks, derived from MediaPipe's canonical FACEMESH_* connection sets and exposed via the new FaceContourType enum and faceContourMeshIndices table. Requires a mesh (standard/full; null in fast); left/right are subject-relative. See the README "Face Contours" section.
• Detection gates: FaceDetector.create() / initialize() accept minScore and minFaceSize (matching Google ML Kit's setMinFaceSize convention), both defaulting to 0.0 (no filtering) and validated to [0.0, 1.0] (out-of-range or NaN throws ArgumentError). Adds Face.widthFraction (visible face width / image width), the value minFaceSize compares against. minScore only tightens results above the detector's internal 0.5 floor. See the README "Detection Gates" section.
• Expose confidence scores: Face.score (detector face-presence confidence) and Face.meshScore / FaceMesh.score (mesh model's confidence, null in fast), plus FaceLandmark.callWithScore(). All from existing outputs, so no added cost. See the README "Detection Score" section.
• Fix: face mesh z is now scaled consistently with x/y (previously left in the model's input-pixel units), making the mesh usable for 3D geometry such as head pose. x/y rendering and iris landmarks are unaffected.
• Overlay helpers (DetectionsPainter, CameraDetectionPainter, FaceDetectionCameraOverlay) gain an opt-in showPoseAndScores flag (default false) drawing a per-face card with confidence and head-pose angles, with toggles in the example app.
• Docs: documented all public enum values and added README sections for the above plus detectFacesWithSegmentation / DetectionWithSegmentationResult.

• Update flutter_litert -> 3.2.0
• Import native-only flutter_litert APIs via package:flutter_litert/native.dart so they resolve under static analysis (flutter_litert 3.2.0 moved InterpreterFactory, IsolateRpcClient, IsolateWorkerBase, and TensorFloat32Views behind the native conditional export). No runtime or API change.
• Default the public entry's conditional export to the web implementation, gating native behind dart.library.io, restoring WASM compatibility (pub.dev WASM-ready). No behavior change on any platform.
• Add package:face_detection_tflite/face_detection_tflite_native.dart, a native-only entry point that re-exports the native implementation (isolate workers, model runners, overlay and UI helpers) for code that runs only on native platforms.

• Performance: when getFaceEmbedding follows detectFacesFromBytes on the same encoded image, the detection isolate now reuses the already-decoded image instead of decoding it a second time (one-entry cache keyed by an exact byte match). Saves a full image decode per detect+embed pair (~16 ms at 12 MP; scales with resolution). No API change, and detection and embedding results are byte-identical. The raw-pixel APIs (detectFacesFromMatBytes, getFaceEmbeddingFromMatBytes) are unaffected; the cache holds at most one decoded frame and is released on dispose.

• Update flutter_litert -> 3.1.1
• Add optional LiteRT Next CompiledModel inference via CompiledModel.fromBufferWithGpuFallback (GPU with automatic CPU fallback); enable with useCompiledModel: true. The default engine remains the Interpreter, so existing code is unchanged.
• Decode camera frames through the shared flutter_litert CameraFrameDecodePlan helper.

• Update flutter_litert -> 2.8.3

• Rename detectFaces -> detectFacesFromBytes for clarity (input is encoded image bytes, vs. raw pixels in detectFacesFromMatBytes); detectFaces is kept as a deprecated alias and will be removed in a future release
• Update flutter_litert -> 2.8.0
• Complete Swift Package Manager migration: example uses CocoaPods only for the optional MLKit comparison benchmark

• Remove unused Darwin podspecs for Dart-only iOS/macOS plugin registration.

• Update flutter_litert -> 2.5.8
• Migrate macOS to Swift Package Manager (CocoaPods no longer required)
• Update camera_desktop -> 1.1.6 in example

• Update flutter_litert -> 2.5.5

• Update flutter_litert to 2.5.3 and camera_desktop to 1.1.4

• Add Web mode GPU fallback
• Add video file processing mode to example
• Update flutter_litert -> 2.5.2

• Update flutter_litert -> 2.5.0

• Update flutter_litert -> 2.4.1

• Simplify and DRY example app, extract utility helpers

• Update documentation

• Update flutter_litert to 2.4.0

• Re-export packYuv420, YuvPlane, YuvLayout, and PackedYuv from flutter_litert so live-camera consumers can reach the helper through the face_detection_tflite barrel without a direct flutter_litert import.
• Update flutter_litert to ^2.2.0
• Add FaceDetector.modelVersion constant so consumers that persist detection results have a stable cache-invalidation key. Bumped on changes that alter detection output (model swaps, threshold or preprocessing changes); unchanged across pure refactors or API additions.
• Rewrite the README's Live Camera and Direct Mat Input sections around packYuv420 so every snippet is a real compilable example (no ghost convertCameraImageToMat, no duplicate segmenter variable, no cv.Mat type annotations requiring an unlisted import).

• Remove FaceDetectorIsolate - FaceDetector is now the single unified class running all inference in a background isolate
• Remove irisOkCount and irisFailCount (were deprecated in 5.1.0)
• FaceDetector() constructor is now public; initialize() replaces the old spawn() factory
• initialize() gains withSegmentation and segmentationConfig parameters
• initializeSegmentation() no longer requires re-spawning the detection isolate
• Add getFaceEmbeddingFromMatBytes to mirror detectFacesFromMatBytes for callers with pre-decoded pixel data
• Improve getFaceEmbeddingFromMat performance by transferring raw pixel bytes to the background isolate instead of re-encoding

• Update flutter_litert to 2.0.13

• Update flutter_litert -> 2.0.12

• Update detectFacesFromMatBytes documentation

• Add detectFacesFromMatBytes to FaceDetector: detects faces from raw pixel data without constructing a cv.Mat on the calling thread (zero-copy transfer via TransferableTypedData)

• FaceDetector now runs all inference in a background isolate automatically, matching FaceDetectorIsolate performance
• dispose() is now Future<void> (was void), existing code compiles but should be awaited
• Deprecate FaceDetectorIsolate: use FaceDetector instead
• Deprecate irisOkCount and irisFailCount (not trackable across isolate boundaries)
• Add detectFacesWithSegmentation and detectFacesWithSegmentationFromMat to FaceDetector

• Update flutter_litert 2.0.10 -> 2.0.11

• Update documentation

• Update flutter_litert 2.0.8 -> 2.0.10

• Add Windows XNNPack delegate support (2-5x inference speedup)
• Update flutter_litert 2.0.6 -> 2.0.8

• Update flutter_litert 2.0.5 -> 2.0.6

• Fix Xcode build warnings by declaring PrivacyInfo.xcprivacy as a resource bundle in iOS and macOS podspecs

• Update camera_desktop 1.0.1 -> 1.0.3
• Use shared Point and BoundingBox from flutter_litert 2.0.0
• Refactor isolate worker to use IsolateWorkerBase from flutter_litert
• Consolidate NMS helpers, extract shared _buildPersonMask and _irisCenterFromPoints
• Deduplicate FaceDetector and FaceDetectorIsolate internals

• Update flutter_litert -> 1.2.0
• Refactor to use flutter_litert shared utilities (InterpreterFactory, PerformanceConfig, generateAnchors)

• Update opencv_dart 2.1.0 -> 2.2.1
• Update flutter_litert 1.0.2 -> 1.0.3

• Update flutter_litert 1.0.1 -> 1.0.2

• Update camera 0.11.3 -> 0.12.0
• Update flutter_litert 0.2.2 -> 1.0.1

• Update flutter_litert to 0.2.2
• Add original model cards for archival and documentation

• Migrate iOS CocoaPods -> Swift Package Manager

Breaking changes:

• Remove all deprecated image package-based APIs across FaceDetector, FaceDetectorIsolate, IsolateWorker, model runners (FaceDetectionModel, FaceLandmark, FaceEmbedding, IrisLandmark, SelfieSegmentation), and helper functions
• Remove image package dependency

• Update flutter_litert to 0.1.12

• Swift Package Manager support
• Windows: remove bundled .dll files, as they are no longer needed as of flutter_litert 0.1.4

• Windows: Custom ops (segmentation) fix
• Fix heap corruption crash when switching between segmentation models

• Migrate from tflite_flutter_custom to flutter_litert

• Fix FaceDetectorIsolate hang on Android during batch face embeddings
• 3-4x performance improvement for FaceDetectorIsolate by eliminating redundant nested isolates
• Models inside worker isolates now invoke TFLite directly instead of routing through nested IsolateInterpreters

• Fix Android build: bump tflite_flutter_custom to 1.2.5 (fixes undefined symbol TfLiteIntArrayCreate linker error)

• Fix bug causing auto-bundling to fail on MacOS

• Update all dependencies to latest version(s)

• Selfie segmentation for background removal and virtual backgrounds
• Uses MediaPipe Selfie Segmentation models (general 256×256, landscape 144×256)

• Performance optimizations: pre-allocated inference buffers, early score filtering (~17× fewer box decodes), parallel multi-face processing

• Fixes #3: bug causing crash on non-XNNPack compatible Android devices

• Face recognition via embeddings, enables comparing faces across images
  • getFaceEmbedding() / getFaceEmbeddings() methods on FaceDetector and FaceDetectorIsolate
  • compareFaces() for cosine similarity, faceDistance() for Euclidean distance
  • Uses MobileFaceNet model (~5MB, ~18ms inference)

• Fix crash on Windows platforms

• Add FaceDetectorIsolate for background thread detection

• Native image processing with opencv_dart for ~2x performance improvement via SIMD acceleration
  • detectFaces() now uses OpenCV internally
  • New detectFacesFromMat() method for camera streams (avoids repeated encode/decode overhead)
• XNNPACK delegate enabled by default for 2-5x CPU speedup (use PerformanceConfig.disabled to opt out)
• Benchmark tests

Breaking changes:

• Replace math.Point<double> type references with Point
• Change face.mesh.isEmpty to face.mesh == null
• Access mesh points via face.mesh?.points[i] or face.mesh?[i]
• Replace face.irises → face.eyes
• Replace IrisPair → EyePair
• Replace iris.center → eye.irisCenter
• Replace iris.contour → eye.irisContour

Improvements:

• Performance and speed improvements
  • Optimize bilinear sampling with direct buffer access, 20-40% speed improvement
  • Fast-path frame registration
  • Parallel iris refinement
  • Isolate-based image-to-tensor conversion.
• Improved test suite, added integration tests

• EyePair class and eye mesh landmarks (71 points per eye)
• Add contour getter for accessing visible eyelid outline (first 15 of 71 points)
• Add eyeLandmarkConnections constant for rendering connected eyelid outline
• Add kMaxEyeLandmark constant defining eyeball contour point count

• Guard iris ROI size and fall back when eye crop collapses