Skip to content
TayHCodePublic

About

Local photo and video search by meaning, with a B-roll finder. Runs on your machine: CPU, CUDA or Apple GPU.

Resources

Contributing

Security policy

Stars

1 star

Watchers

0 watching

Forks

Repository files navigation

shotseek

shotseek is a small local tool for people with a lot of media: a job's asset folder, a creator's camera roll, or a club's shared drive. Point it at a folder of photos and videos. It builds a search index on your computer. Ask your own Claude or Codex agent for pictures or video moments, or hand it a talking-head video and get B-roll suggestions for each passage. It runs offline after a one-time model download, uses no paid service, and never changes your media.

What you can ask for

  • "Find a landscape picture of latte art for the cafe website."
  • "Find a portrait photo of our hiking club at a trailhead for a social post."
  • "Find B-roll for this product shoot A-roll, especially hands unboxing and close-ups of the ceramic mug."

Requirements

Python 3.11+ (3.12 was used for the platform checks), ffmpeg and ffprobe, and space for packages, model weights, package caches, and your index. Linux, Apple Silicon macOS, and native Windows have been tested with the qualifications below. Additional models need more space. See platform setup for isolated Python installs, portable ffmpeg, and Windows setup. Native Windows x64 also needs the Microsoft Visual C++ v14 Redistributable for the Python packages that load native libraries. Installing it may require administrator approval and a restart.

shotseek has been tested on a desktop with an NVIDIA RTX-class GPU. It runs on CPU by default and does not need a GPU; add --device cuda to use one.

Quickstart (about ten minutes)

About ten minutes to a first search means searching a partial index. Package and model download time depends on the connection; a large or cloud-mounted library can take hours to finish indexing.

  1. Without Git, use Code → Download ZIP on this repository, extract it, and open a terminal in the extracted folder containing pyproject.toml. An existing source archive works too. With Git, clone this repository:

    git clone <this repository's address> shotseek && cd shotseek
  2. Install uv and let it create a Python 3.12 environment. It downloads Python when needed, avoiding old macOS Python and Linux's missing ensurepip. On Linux/macOS:

    curl -LsSf https://astral.sh/uv/install.sh | sh
    export PATH="$HOME/.local/bin:$PATH"
    uv venv --python 3.12
    . .venv/bin/activate

    The installer normally writes uv and uvx to ~/.local/bin and updates your shell profile. The PATH line makes them available in this shell; opening a new terminal also works. If XDG_BIN_HOME, UV_INSTALL_DIR, or UV_UNMANAGED_INSTALL selects another directory, replace $HOME/.local/bin in that line with the chosen directory. A custom or unmanaged install may provide no env setup file. For an isolated install without profile changes, see platform setup.

    On native Windows, use PowerShell:

    winget install --id astral-sh.uv -e

    Reopen PowerShell so uv is on PATH, then return to the checkout. Alternatively, use the PowerShell installer in the current session:

    New-Item -ItemType Directory -Force .scratch | Out-Null
    Invoke-WebRequest -UseBasicParsing https://astral.sh/uv/install.ps1 -OutFile .scratch\install-uv.ps1
    & "$env:SystemRoot\System32\WindowsPowerShell\v1.0\powershell.exe" -NoProfile -ExecutionPolicy Bypass -File .scratch\install-uv.ps1
    $env:PATH = "$env:USERPROFILE\.local\bin;$env:PATH"
    uv venv --python 3.12

    The PowerShell installer normally writes uv.exe and uvx.exe into %USERPROFILE%\.local\bin and updates your user PATH. Restart PowerShell or use the explicit PATH line above to make it available in this session. Follow the printed location if you chose another install directory.

    After winget, also run uv venv --python 3.12. No venv activation or execution policy change is needed to run the venv executables. The installer above uses Bypass only in its child process; it leaves the caller and persistent policies unchanged. An enforced Group Policy can still block it. In subsequent commands replace shotseek with & .\.venv\Scripts\shotseek.exe.

    Alternative without uv: first run python3 --version; it must be 3.11+. If it is older, use uv above, which downloads a compatible Python. Then run python3 -m venv .venv && . .venv/bin/activate. On Debian/Ubuntu, an ensurepip/venv failure needs sudo apt install python3-venv (or the matching version package, such as python3.12-venv); use uv if you cannot install it. Windows with Python 3.12 already installed can use py -3.12 -m venv .venv. For this alternative replace uv pip install below with python -m pip install (Windows: & .\.venv\Scripts\python.exe -m pip install).

  3. Install both ffmpeg and ffprobe before indexing. They must be on PATH; Python dependency wheels do not provide these command-line programs. Debian/Ubuntu: sudo apt install ffmpeg. macOS with Homebrew: brew install ffmpeg. Windows: winget install --id Gyan.FFmpeg -e, then reopen PowerShell and return to the checkout.

    Without administrator access, Linux x86_64 can use a John Van Sickle static build:

    mkdir -p .scratch/ffmpeg
    curl -L https://johnvansickle.com/ffmpeg/releases/ffmpeg-release-amd64-static.tar.xz -o .scratch/ffmpeg.tar.xz
    tar -xJf .scratch/ffmpeg.tar.xz -C .scratch/ffmpeg --strip-components=1
    export PATH="$PWD/.scratch/ffmpeg:$PATH"

    On Apple Silicon without Homebrew, download both arm64 executables from OSXExperts:

    mkdir -p .scratch/ffmpeg
    curl -L https://www.osxexperts.net/ffmpeg9arm.zip -o .scratch/ffmpeg.zip
    curl -L https://www.osxexperts.net/ffprobe9arm.zip -o .scratch/ffprobe.zip
    unzip .scratch/ffmpeg.zip -d .scratch/ffmpeg
    unzip .scratch/ffprobe.zip -d .scratch/ffmpeg
    chmod +x .scratch/ffmpeg/ffmpeg .scratch/ffmpeg/ffprobe
    export PATH="$PWD/.scratch/ffmpeg:$PATH"

    Intel Macs need Intel builds from that page. A quarantined browser download may need approval for that individual file in macOS Privacy & Security. On Windows without winget, use Gyan's portable essentials ZIP:

    New-Item -ItemType Directory -Force .scratch | Out-Null
    Invoke-WebRequest -UseBasicParsing https://www.gyan.dev/ffmpeg/builds/ffmpeg-release-essentials.zip -OutFile .scratch\ffmpeg.zip
    Expand-Archive .scratch\ffmpeg.zip -DestinationPath .scratch\ffmpeg
    $ffmpegBin = (Get-ChildItem .scratch\ffmpeg -Directory | Select-Object -First 1).FullName + '\bin'
    $env:PATH = "$ffmpegBin;$env:PATH"

    These portable PATH changes last only for the current shell. Verify both:

    ffmpeg -version
    ffprobe -version
  4. On Windows x64, install the Microsoft Visual C++ v14 Redistributable before loading PyTorch or the speech runtime:

    New-Item -ItemType Directory -Force .scratch | Out-Null
    Invoke-WebRequest -UseBasicParsing https://aka.ms/vc14/vc_redist.x64.exe -OutFile .scratch\vc_redist.x64.exe
    Start-Process .scratch\vc_redist.x64.exe -ArgumentList '/install' -Wait

    Complete the installer and restart if requested. Stop if installation fails; reinstalling Python wheels does not install the Windows native runtime.

    On Linux or Windows CPU installs, install PyTorch and torchvision together from the CPU wheel index. Both must use matching builds. Skip this first command on macOS; its PyPI wheels support CPU and Apple Silicon MPS. Windows CUDA users should use the matching CUDA pair in platform setup instead (about 2.75 GB for torch alone).

    uv pip install 'torch>=2.2' 'torchvision>=0.17' --index-url https://download.pytorch.org/whl/cpu
    uv pip install .

    The second command installs shotseek and remaining dependencies from PyPI; do not keep the CPU wheel index as a global index. CPU torch wheels measured roughly 125–200 MB, plus other dependencies; a Linux CPU environment measured about 1.6 GB. Sizes vary by version and platform.

    For the Python venv alternative in step 2, the equivalent commands are:

    python -m pip install 'torch>=2.2' 'torchvision>=0.17' --index-url https://download.pytorch.org/whl/cpu
    python -m pip install .

    On macOS, skip the first command as above. On Windows, check the native libraries before downloading model weights:

    & .\.venv\Scripts\python.exe -c "import torch, torchvision, open_clip, faster_whisper; print(torch.ones(1).item())"

    Expect 1.0. A DLL-loading error means the native runtime or matching wheels still need attention. Continue only after the imports succeed.

  5. Check local requirements, including torchvision's CPU operators. A missing config is a warning at this stage. If ffmpeg or ffprobe is absent, JSON shows a fail check with detail Missing and hint Install ffmpeg and ffprobe., ok: false, and exit 6. Install them using step 3, restore PATH in this shell, and rerun doctor:

    shotseek doctor
  6. Choose a media folder; the index stays outside it:

    shotseek init --library media=/path/to/media

    Quote the entire argument when the path contains spaces. On native Windows:

    & .\.venv\Scripts\shotseek.exe init --library "media=C:\Media\Generated Photos"

    To add another folder later, keep the existing config and use two separate arguments (no =):

    shotseek library add extra "/path/to/second media folder"
    shotseek library list

    Then run shotseek index to index both libraries. Do not rerun init to add a folder: init --force replaces the whole config, including libraries and custom settings. Use separate --config and --index-dir paths when sharing an account with another installation.

  7. Start a five-minute indexing pass. The first pass downloads the image model (stored once in the local model cache):

    shotseek index --time-budget 300
  8. Search the committed partial index:

    shotseek search "a barista pouring latte art" --format text
  9. Finish indexing. This command can be stopped and resumed:

    shotseek index

Downloaded default model files measured 605.14 MB for OpenCLIP image/text and 147.88 MB for base speech (decimal MB). Package and model disk use depends on installed versions. The measured PyPI environment occupied 6.34 GB and its pip cache 3.23 GB, including CUDA runtime dependencies; execution used CPU. Leave additional space for the media index (all sizes here use decimal units).

Connect your agent

For Claude Code, copy agents/claude/shotseek/ to ~/.claude/skills/shotseek/. For Codex, paste agents/codex/AGENTS.md into a project's AGENTS.md or ~/.codex/AGENTS.md. For another agent, paste agents/generic/PROMPT.md into its instructions. See agent setup. Then say: "Use shotseek to find a landscape cafe photo for the website."

B-roll for a talking-head video

shotseek broll --media any --point "coffee and pastry at the counter" --point "opening the cafe at dawn"
shotseek broll --aroll /path/to/aroll/intro.mp4 --save-transcript /path/outside/library/aroll.json --format json --output /path/outside/library/beats.json
shotseek broll --beats /path/outside/library/beats.json --format csv --output /path/outside/library/broll.csv

B-roll searches videos only by default; add --media any to include photos, as the first example does, or --media photo for stills only. A-roll must contain speech; silent videos need --transcript or --point instead.

The --aroll command above writes the complete editable B-roll document to beats.json with --output; --save-transcript writes the separate transcript aroll.json, which is not the beats file. Replace the three example paths with your A-roll and an output folder outside your libraries. The two commands run in order as written, including without edits. To refine the plan, open beats.json between runs and add two to four concrete visual_queries strings to weak beats, then run the --beats command to write broll.csv.

You can also supply --transcript with text, SRT, VTT, or JSON. Abstract speech is often a poor description of visible footage. Each beat recommends one candidate, preferring variety. The default local speech model is base. Automatic planning uses consecutive 4–10 second windows, the need score, and five generic topic hints. Every beat includes need_score, need_reasons, topic, suggested_broll, search_terms, and a 3–6 second placement. Low-need, short, filler, and over-limit beats stay visible with a skip reason. Points and explicit visual queries are always searched.

Use --segmenter pauses for pause and similarity boundaries. --min-need 0.45 and --max-beats 12 are defaults; --max-beats 0 removes the limit. --include-filler searches every beat for that invocation. Keep the source, revision, and IDs in an edited document so source exclusions survive reruns. --keep-near-duplicates disables visual similarity exclusion while retaining physical identity and content-sample checks.

Recommendations rank the available candidates; they can be unrelated when the library has no matching footage. Review each choice before using it. The text score is cosine similarity followed by (relevance), a z-score relative to the filtered candidates, not a probability or confidence percentage. With only two candidates, relevance can be +1 and -1 even when both are poor matches. B-roll ranking is only meaningful on real footage; generated test clips can rank arbitrarily.

Using a GPU

Install matching CUDA builds of both PyTorch and torchvision using the PyTorch selector and your driver version (the tested Windows torch wheel was about 2.75 GB); platform setup records the tested Windows pair. Then add --device cuda or set device = "cuda" under [model] in the config. cuda:N, mps, and auto are also accepted by image embedding. CPU is the default, and search uses CPU unless a flag or environment variable selects another device. Larger OpenCLIP models can improve visual matching: ViT-L-14 with laion2b_s32b_b82k (768 dimensions) is the suggested GPU choice. The default is ViT-B-32 with laion2b_s34b_b79k (512 dimensions). large-v3 is an optional larger speech model. Speech on CUDA needs CUDA libraries; initialization failure falls back to CPU with a warning. Local speech does not support MPS.

Where your media can live

Use local folders, a NAS path, or a mounted cloud folder such as Google Drive. There is no cloud API integration. Keep the index on local storage. See storage guidance for mounts, remapping, and safe pruning.

What it stores and what it never does

On Linux and macOS, config defaults to ~/.config/shotseek/config.toml and the index to ~/.local/share/shotseek/index; XDG settings and explicit flags override them. Native Windows uses %APPDATA%\shotseek\config.toml and %LOCALAPPDATA%\shotseek\index. The index contains relative file paths, metadata, vectors, transcripts, and 320 px JPEG thumbnails. It is as private as the library: protect it accordingly. Set thumbnails = false under [index] to disable future thumbnail storage. A rebuild removes existing thumbnails while preserving cached transcripts.

Each 512-dimensional float16 vector occupies 1,024 bytes, plus metadata and optional thumbnails. Thumbnail sizes depend on image content. Generated-input measurements describe only the tested fixtures, not real media or mounts.

CPU measurements on Linux used an eight-core desktop CPU, eight PyTorch threads, and generated inputs, with models already cached. The PyPI installation included CUDA runtime packages, but all model execution below used CPU. These fixtures cannot predict performance or match quality on a real library.

Generated-input operation Wall time
Index 300 solid-colour JPEG photos, model already loaded 12.639 s
Add 12 frames from one 60-second video 1.322 s
First text searches, including image/text model loading 0.976–1.124 s
Repeat searches using cached query vectors 0.0049–0.0058 s
Transcribe 4.465 seconds of synthetic speech with base 2.297 s
Reuse the cached transcript 0.0357 s

The cached image model loaded in 2.018 seconds. Generated JPEG thumbnails averaged 1,829 bytes; detailed photographs can produce larger thumbnails. A separate 400,000-vector synthetic search took 1.255 seconds with 523 MB peak resident memory. On four generated images, the 448-pixel pre-fit changed cosine similarity by at most 0.00337; direct image/text model calls agreed within 0.0001.

The tool never writes to a library, copies original media, reads GPS data, phones home, checks for updates, or calls a model API. It makes no network calls of its own; model libraries download weights once when first needed. Model loading and dependency download progress appear on stderr. A Hugging Face unauthenticated-request warning is informational; no account is required. --offline or SHOTSEEK_OFFLINE=1 blocks them, and missing cached weights produce a clear error. Temporary extracted A-roll audio is deleted after transcription.

Commands

Command Purpose
init Create configuration with named libraries
library library add NAME PATH, library list, or library remove NAME
index Incrementally index photos and video frames
status Inspect progress and index coverage
search Rank photos and video moments
broll Find candidates for each spoken passage or point
transcribe Transcribe locally without an image index
show Inspect one file and all its frames
doctor Check installation and storage; --report adds a pasteable platform report
compact Reclaim dead vector rows
schema Print JSON contract schemas
agent-prompt Print agent instructions

Use --config /path/to/config.toml and --index-dir /path/to/index to keep separate configurations and indexes outside your media folders.

Default output is JSON; search and B-roll also support CSV and text. search --raw-frames returns individual frames in descending score order, without moment merging, per-file limits, or duplicate folding. Batch query CSVs use query_id,query; see invented queries.

index can exit 0 even when individual files fail: successful files remain indexed, and warnings, totals.failed, and failures_sample report failures. Use shotseek index --strict for exit 8 when files fail, or shotseek index --retry-failed --strict to retry them. Storage and other fatal errors still have their own nonzero codes. Files not reached because a folder became unavailable are counted separately in totals.not_reached and each library's not_reached, with a files_not_reached warning. Reconnect and rerun indexing normally; these files are not recorded as failed. The run still exits 4 and preserves completed rows. Put command-specific flags such as --device after the subcommand: shotseek search "a red mug" --device cpu.

index --max-files-per-library N caps newly indexed files separately for each library and leaves a resumable partial run. There is no default file cap.

Under [indexing], frame_reader accepts auto, opencv, or ffmpeg. auto uses the guarded OpenCV seek/read path with an ffmpeg fallback. video_sampling = "interval" samples every five seconds by default, spreading capped samples across the full sample range. video_sampling = "fixed" uses video_samples_per_file = 4 positions from 15% through 85% of the video. See the output reference for fields, flags, and exit codes.

Limitations and not in this version

English queries only. Video timestamps are accurate to the sampling interval. RAW files are counted but not decoded: portable generated RAW fixtures are unavailable, files are large on slow mounts, and JPEG or HEIC companions are common. HEIC requires pillow-heif. Animated images use their first frame.

There is no graphical or web interface, MCP server, cloud API, OCR, library speech search, shot-boundary detection, clip export or copying, file watcher, background service, timer, or package-registry publication in this version. Face recognition and identifying people are never planned. Platform checks use small generated fixtures and do not establish performance on large libraries. Mounted-storage checks cover small generated libraries and available media; cold cloud placeholders, completed cloud synchronization, physical network partitions, packet-level throttling, and large network libraries remain unverified. A-roll transcription does not make library speech searchable.

On macOS, OpenCV and PyAV wheels can emit duplicate AVFoundation class warnings (AVFFrameReceiver / AVFAudioReceiver). The tested CPU/MPS runs completed without a crash; this does not establish that every wheel combination is safe. See platform setup if these warnings accompany a failure.

Known issues

A few timing-based tests may flake on a heavily loaded machine. One unexplained, one-off Python 3.12 test-step failure during release verification did not reproduce in two clean runs.

What has been verified

Platform checks on 2026-10-09 used the default OpenCLIP model, generated media, Python 3.12, isolated package/model folders, and no administrator changes. Earlier checks identified prerequisite and shell gaps now covered in the quickstart and platform setup.

Platform or storage Verified coverage Still untested
Linux CPU Quickstart, doctor report, 60 generated photos + 3 short videos, image/video search, long paths, compact and rebuild Large real libraries
Apple Silicon macOS, CPU and MPS Quickstart, doctor reports, the same media and searches on both devices; CPU ffmpeg paths up to 395 characters, compact and rebuild; doctor identifies the CPU MPS speech (unsupported); browser-download Gatekeeper approval
Native Windows, CPU and NVIDIA RTX-class CUDA PowerShell quickstart, doctor reports, generated photos/videos on both devices; UTF-8 legacy pipes, 376-character media paths with long-path policy disabled, auto/ffmpeg decoders, compact and rebuild AMD GPU execution, CUDA speech
NAS, read-only SSHFS Two bundled OS icons indexed and searched through a live remote mount; unchanged rerun; source hashes and timestamps unchanged Personal or large NAS libraries, mount outages

Mounted storage

Live checks on 2026-10-10 used generated media, local indexes, and the default CPU OpenCLIP model. A verified read means the file opened normally through its mounted path; it does not establish an online-only cloud state.

Storage and platform Verified live coverage Mounted / identical local index time
NFS 4.1, Linux 64 photos and three videos; init/add, index/search, unchanged and mutation reruns, doctor, point/spoken B-roll, permission failure/retry, writer locking; bounded injected listing stall 9.848 / 9.839 s
Google Drive for desktop, macOS File Provider The tested install read 64 photos and three videos normally; 424-character paths, case/Unicode, index/search/reruns, doctor, spoken B-roll, permission failure/retry, writer locking; provider reported a generated file uploading 8.651 / 8.042 s
Google Drive for desktop, Windows streaming Available files read normally: 64 photos and three videos, 366-character paths with long-path policy disabled, case/Unicode, index/search/reruns, doctor, spoken B-roll, partial-write failure/retry, writer locking 8.719 / 8.703 s
SMB, macOS 64 photos and three videos; index/search, second library, doctor, spaces/Unicode, unchanged pass with zero media reads, case-alias overlap/write guards, share-removal failure and recovery 8.456 / 4.043 s
SMB, Linux GVfs/gio The same generated-media flow, zero-read unchanged pass, share-removal recovery and case-alias guards despite synthetic inode IDs 5.586 / 5.095 s
SMB, Windows drive letter The same generated-media flow, zero-read unchanged pass, share-removal recovery and case-alias guards 4.907 / 4.688 s
SMB, Windows UNC The same generated-media flow and guards; existing-index search also passed while the scratch share was absent 5.141 / 4.688 s

Timings are single whole CLI commands with models cached, including model startup; SMB timings use matching cached rebuilds. They do not establish large-library throughput or ranking quality. The earlier platform timings below measure a different, in-process flow.

Fault checks retained searchable successes, pruned no rows after share removal, and recovered on retry. Partial writes and injected slow reads were reported; wrong credentials and temporary account expiry rejected fresh SMB connections. Case aliases cannot bypass library-overlap or output-write guards, and writes are refused when alias verification is incomplete. Account expiry does not test password-policy expiry.

Not verified: truly online-only (cold) Drive files, Drive mirrored mode, physical network partition or packet-level throttling, password-policy expiry, guaranteed offline availability, completed cloud upload, large real libraries, or other SMB servers/dialects. Tests could not produce and independently confirm a cold Drive file. The later exact scratch folder was absent on the Mac install, so that supplemental cold-file check could not run there. The Windows Drive ACL-denial attempt also did not establish native permission-denial coverage.

An additional Windows streaming check used four cloud-created JPEGs of 995–1,175 bytes. First reads succeeded in 172–231 ms each, with one content open per file; search and doctor passed, and unchanged indexing opened no media. Windows reported only Normal attributes, without offline/recall/pinned/reparse flags. Unpinning left those attributes unchanged; rebuild reads took 2–5 ms. Neither cloud creation nor these timings proved a first or second cold pass. Doctor checks reachability, not individual cloud placeholder state.

For online-only files, shotseek opens the mounted path and the Drive client triggers a download. Slow or failed reads are retried and reported as timeout or read_error; successful files remain searchable. Let downloads finish and rerun with --retry-failed --strict. A bounded worker can stop the CLI waiting, but cannot cancel the provider's download or a stalled kernel read. Keep the model cache, config, and index on local storage outside the mounted library. See network drives and cloud mounts for setup, retry limits, change-detection limits, and storage-specific gaps.

The measured indexing passes used 60 generated photos and three six-second videos, with the model already loaded and eight CPU threads. Each device used its default batch size (16 on CPU, 64 on GPU); media decoding still used CPU. These are single small-fixture passes, not reliable GPU speedup estimates.

Platform and device Files Index time Files/s
Linux CPU, eight-core desktop 63 7.574 s 8.318
Apple Silicon macOS, CPU 63 0.901 s 69.949
Apple Silicon macOS, MPS 63 0.862 s 73.121
Windows CPU 63 1.223 s 51.522
Windows NVIDIA RTX-class GPU, CUDA 63 0.637 s 98.907
Second Linux machine, CPU 63 6.438 s 9.786
NAS read-only SSHFS, CPU (two OS icons) 2 0.280 s 7.134

The maintainer previously verified a run limited to 200 videos. Libraries of tens of thousands of files remain unverified. All timings here are generated fixtures or the explicitly identified NAS sample, not a real-media benchmark.

To help close these gaps, run this on your machine and paste the result into an issue:

shotseek doctor --report

On tested macOS wheels, the two AVFoundation duplicate-class warnings (AVFFrameReceiver / AVFAudioReceiver) on stderr were harmless and the runs completed. They can also appear above a pasted report if stderr is merged; shotseek doctor --report > doctor-report.txt saves just the report on stdout.

It prints your OS, Python, CPU, GPU/MPS availability, whether ffmpeg and ffprobe are found, whether the model is cached, and PASS or FAIL for indexing and searching three generated photos in a scratch folder. The first run downloads the model if it is not cached; add --offline to forbid that and --device cuda (or mps) to test an accelerator. Home folders are shown as ~ and your media and config paths are not printed. To test a NAS or mounted folder, run shotseek init --library media=<path> and shotseek doctor first; doctor then also reports whether each library is reachable.

Related tools

Full photo managers with smart search include Immich and PhotoPrism; shotseek is a small command-line index for agents.

Third-party components and licenses

OpenCLIP and its default weights, faster-whisper and its models, PyTorch, torchvision, OpenCV, Pillow, pillow-heif, and numpy each have their own terms. THIRD-PARTY.md lists them, the model licenses, and the native libraries that some wheels bundle. ffmpeg is an external program you install yourself, not bundled.

Some dependency wheels bundle GPL/LGPL native libraries, including GPL codecs in pillow-heif and PyAV. shotseek does not redistribute these wheels. Review their terms before redistributing an installed environment or application.

License

MIT. See LICENSE.

About

Local photo and video search by meaning, with a B-roll finder. Runs on your machine: CPU, CUDA or Apple GPU.

Resources

Contributing

Security policy

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Used by

Contributors

Languages