v1.10.2

CloakFrame is the new name for Redactly. This release applies the new identity
throughout the app while preserving existing user data.

Improvements

• Renames the app, executables, packages, namespaces, build targets, assets, and
  translations to CloakFrame
• Imports settings from earlier Redactly releases on first launch
• Reuses face and license-plate models already downloaded by earlier releases

v1.10.1

Makes very large video processing practical by keeping all temporary data on
the destination drive.

Improvements

• The private source snapshot and the encoder working file are now stored in
  hidden folders inside the chosen output folder instead of the system
  temporary directory, so processing a large video only needs free space on
  the destination drive
• Finished videos are published with an instant atomic rename instead of a
  full copy when the temporary location and the destination were on
  different drives, removing tens of gigabytes of extra writes for long 4K
  footage
• Linux systems with a RAM-backed /tmp no longer run out of space when
  processing videos larger than available memory

All temporary folders are removed automatically when a run completes, fails,
or is cancelled, and the snapshot's tamper protection is unchanged.

Platform notes

• macOS 15 (Sequoia) is supported again; the macOS 26 requirement introduced
  in v1.10.0 has been lifted

v1.10.0

Hardens how Redactly reads and writes files end-to-end and scales processing
up for very large photos and videos.

Security

• Output files are published atomically inside the chosen output folder using
  symlink-safe, no-replace writes on every platform, so crafted links or files
  swapped mid-run can never redirect or overwrite results outside it
• Videos are processed from a private snapshot of the source, so edits or
  swaps to the original while a run is in progress cannot affect the output
• Built-in detection models are verified against a pinned SHA-256 digest
  before they are loaded, and custom ONNX models get stricter shape and type
  validation
• Image containers (TIFF, PNG, WebP, animated formats) are parsed with strict
  bounds and overflow checks before decoding

Improvements

• Large media limits are now generous: input files up to 2 GB and images up
  to 512 megapixels are accepted, and the image pipeline adapts its
  parallelism to available memory instead of using a fixed thread count
• Crowded scenes track more reliably: up to 1024 detections per frame and
  65,536 tracks per video are supported, and license plate detection handles
  denser candidate sets
• A single unreadable file in a batch no longer slows down processing of the
  remaining items
• Loaded detection models are reliably reused between runs, including on
  systems where GPU acceleration falls back to the CPU, making repeat runs
  start much faster
• Detection results with invalid coordinates are filtered out before
  matching, fixing rare inconsistent masking on corrupted frames
• A face seen clearly in just a single frame is now kept and masked instead
  of being discarded as noise
• Saving to drives that do not support atomic no-replace renames (exFAT or
  FAT32 USB drives, SD cards, and some network shares) now works on macOS

Platform notes

• macOS builds now require macOS 26 or newer
• Windows CI and release builds set up MSVC directly without third-party
  actions

v1.9.1

Adds GPU-accelerated detection for Linux source builds and refreshes the
official build stack across all supported platforms.

Improvements

• Linux source builds can now use CUDA on NVIDIA GPUs or MIGraphX on supported
  AMD GPUs, with automatic CPU fallback when a provider is unavailable or
  rejects a model
• GPU detection still follows the existing acceleration setting, so it can be
  disabled without rebuilding the application
• Official builds now use newer, aligned Qt, OpenCV, and ONNX Runtime versions,
  with updated Linux, Windows, and macOS build environments
• Build requirements are checked explicitly: Qt 6.8.1 or newer and OpenCV
  4.10.0 or newer are required

Linux packaging note

• The official AppImage continues to use CPU inference. CUDA or MIGraphX
  detection requires a source build linked against a GPU-enabled ONNX Runtime

v1.9.0

Adds video review and more anonymization choices, strengthens tracking and
masking, and expands Redactly's language and video-output options.

New features

• Review detected video tracks on a timeline before encoding and exclude false
  tracks from the entire output
• Use the new built-in smiley sticker anonymization style
• Choose H.264 for broad compatibility or HEVC for smaller video files
• Use Redactly in Japanese; Japanese systems select it automatically on first
  launch, alongside the existing English and Korean options

Improvements

• Video tracks are matched more safely so masks are less likely to jump to an
  unrelated face
• Soft mask edges now blend through the padding area while keeping the detected
  region fully covered
• Korean interface wording has been polished, and the README has been
  consolidated into a single English document

v1.8.1

English

Video processing is faster on macOS, and batch runs now protect existing
output files and report incomplete results more clearly.

Improvements

• Face detection models now use static input dimensions so CoreML can run the
  whole graph instead of leaving unsupported work on the CPU
• Video output uses Apple's VideoToolbox hardware H.264 encoder when available,
  with automatic fallback to software encoding
• Redactly checks every planned output before processing and refuses to start
  if a file would be overwritten or two inputs would produce the same path
• Runs with failures, skipped files, or outputs containing no redacted regions
  now finish as Review required with a clearer summary
• Very low detection thresholds no longer promote weak video detections into
  thousands of false tracks that obscure unrelated parts of the frame
• Update notifications now show the release notes in the app's selected
  language, with a choice to update now or postpone

한국어

macOS에서 동영상 처리 속도가 향상되었으며, 이제 일괄 처리 시 기존 출력
파일을 보호하고 완료되지 않은 작업 결과를 더욱 명확하게 알려줍니다.

개선 사항

• 얼굴 탐지 모델에 고정 입력 크기를 적용하여 CoreML이 일부 작업을 CPU에
  맡기지 않고 전체 그래프를 실행할 수 있도록 개선했습니다
• 가능한 경우 Apple의 VideoToolbox 하드웨어 H.264 인코더를 사용하며,
  사용할 수 없으면 소프트웨어 인코딩으로 자동 전환됩니다
• 처리 전에 생성할 모든 출력 경로를 확인하며, 기존 파일을 덮어쓰거나 두
  입력 파일이 같은 경로에 저장될 경우 작업을 시작하지 않습니다
• 실패하거나 건너뛴 파일 또는 가림 처리된 영역이 없는 출력물이 있으면
  작업 상태를 검토 필요로 표시하고 더욱 명확한 요약을 제공합니다
• 탐지 임계값이 매우 낮더라도 약한 동영상 탐지가 수천 개의 잘못된 트랙으로
  승격되어 화면의 무관한 영역을 가리지 않도록 개선했습니다
• 업데이트 알림에서 앱 설정 언어에 맞는 릴리스 노트를 표시하고, 바로
  업데이트하거나 나중으로 미룰 수 있습니다

v1.8.0

Video face detection is more accurate: it now analyzes video at the full
detection resolution, so it catches faces it used to miss and stops covering
signage it mistook for a face.

Improvements

• The fast face model now analyzes video at the full detection resolution
  instead of a third of it, so smaller and partly turned faces that used to
  slip through are covered
• Fewer false masks: a logo or sign that briefly resembles a face no longer
  stays covered for the rest of the shot
• Masks follow faces more faithfully — they no longer glide into place a few
  frames early, and stay on a face through brief detection dropouts

Details

• The bundled fast model ships with a fixed 640 px input; it is now adapted in
  memory at load time to run at the requested video resolution, with a probe
  check and automatic fallback to its native size, or to the CPU, if a GPU
  backend can't run it
• A track needs several confident detections to be kept, so sparse false
  positives are dropped — while a short but clearly detected face is still
  covered
• A track that goes too long without a confident detection ends instead of
  coasting indefinitely, and moving faces get more leeway than static signage

v1.7.1

Fixes video redaction failing on macOS with GPU acceleration, and stops a
redaction mask from sliding into place when a new face appears.

Fixes

• Video redaction on macOS no longer stops with a CoreML error partway
  through: the face and license-plate models that Apple's GPU backend can't
  run now fall back to the CPU automatically, while the models it can run
  stay on the GPU
• A newly appeared face is now covered where it appears, instead of the mask
  sliding in from where a different face left the frame

Details

• Each detector is exercised once as it loads; if the GPU backend rejects it,
  that detector is rebuilt on the CPU. Photo detection, the fast 640 px face
  model, and the Windows/Linux GPU paths are unaffected
• Detection that falls back to the CPU now runs across all cores instead of a
  single thread, cutting the accurate face model's analysis time on
  high-resolution video by roughly two-thirds
• A track that loses its face no longer drifts across the frame to grab an
  unrelated new one, and a gap that would imply an implausibly fast jump is
  left uninterpolated rather than drawn as a slide

v1.7.0

Video redaction is dramatically faster, and mosaics now fully cover close-up
faces.

Improvements

• Video processing is several times faster — a 3-minute 1080p clip that used to
  take over ten minutes now finishes in under a minute
• Blur redaction is no longer the bottleneck: it renders at a fraction of the
  previous cost with the same visual strength
• Mosaics on large, close-up faces are now properly coarse, so a face stays
  unrecognizable no matter how much of the frame it fills

Details

• Blur now runs on a downscaled copy of each region and is scaled back up,
  capping the kernel cost while preserving the same blur strength
• The soft-edge feather mask is computed once per face size and reused across
  frames regardless of position, instead of being rebuilt for every frame
• The mosaic block count is capped so a face is always reduced to at most a
  dozen cells across, keeping large faces obscured
• Frame redaction runs in parallel across CPU cores, and encode writes overlap
  the masking work, so both passes use the machine more fully

v1.6.1

Video tracking is now scene-cut aware, fixing mosaic ghosting in fast-cut
footage.

Fixes

• Mosaics no longer linger for up to a second after a hard cut, drifting
  across the new shot
• A mosaic can no longer jump to a different person who appears in a
  similar position right after a cut

Details

• The analysis pass now detects shot boundaries, and tracks stop cleanly at
  every cut: no track survives a scene change, no gap is interpolated
  across one, and track ends are never extended past one — in both tracking
  directions
• A cut is only declared when the frame-to-frame change stands well above
  the recent motion level and persists for the following frames, so
  camera flashes, strobes, and fast pans within a shot do not split tracks
• Coverage inside a shot — through motion blur, side profiles, and brief
  occlusions — is unchanged
• The log reports how many scene cuts each video contained