SiftVid
Attribution

Licensing

SiftVid itself is proprietary software owned by SiftVid. It stands on open-source work by other people. This page states what you get when you buy SiftVid, names every third-party component, gives the license each one is under, and explains what those licenses do and do not require of you.

Effective: 6 August 2026Version: 1.0Publisher: SiftVidContact: reply@siftvid.com

01Scope

Draft - counsel review pending

The FFmpeg and libvips positions below follow the reading that the VLC and HandBrake ecosystems have relied on for years. They are a good-faith engineering record, not legal advice, and have not yet been signed off by counsel.

SiftVid itself is proprietary software published bySiftVid. The components listed on this page are not; each one remains under its own license, held by its own authors.

This page mirrors THIRD-PARTY-NOTICES.md, which ships inside the application and is also readable in the app under About → Open-source licenses. That file is regenerated from the actual dependency tree every time a release is built, so the list in Section 07 always matches what a given installer contains.

02What your purchase gives you

Buying SiftVid buys you a license to use the application, granted by SiftVid on the terms set out in the Terms of Service. It is a one-time purchase, it does not expire, and it includes future updates. It is not a transfer of ownership and it is not an open-source grant.

SiftVid is proprietary and closed-source

SiftVid is not released under any open-source license. Its source code is not published, and no license on this page entitles you to it. You may install and use the software on the devices you personally use; you may not redistribute, resell, sublicense, or modify it, or reverse-engineer it except where the law says you may regardless. The open-source licenses described below attach to the individual components listed — never to SiftVid itself.

That distinction matters because two of those components are under copyleft licenses (GPL-3.0-or-later and LGPL-3.0-or-later). Sections 03 and 04 set out why those licenses reach the components and stop there, which is the same position the VLC and HandBrake ecosystems have taken for years.

03FFmpeg

SiftVid uses FFmpeg to read video metadata, generate thumbnails, transcode for playback, and render exports. SiftVid does not bundle FFmpeg. On first run the application downloads a prebuilt Windows binary from this exact address, published by the BtbN/FFmpeg-Builds project:

https://github.com/BtbN/FFmpeg-Builds/releases/download/latest/ffmpeg-master-latest-win64-gpl.zip

Build variant and license. This is BtbN's GPL variant, not the LGPL one. That is not inferred from the file name: the binary reports its own build configuration, and it contains --enable-gpl together with --enable-version3, with libx264 and libx265 enabled and libfdk-aac disabled. Under FFmpeg's own licensing terms, enabling the GPL-only libraries makes the build GPL, and --enable-version3 elects version 3 of the license. The effective license of the binary SiftVid downloads is therefore the GNU General Public License, version 3 or later. Because libfdk-aac is disabled, this is not BtbN's "nonfree" variant, and the binary is freely redistributable.

Source code offer. The complete corresponding source for those binaries is available at:

  • FFmpeg source: github.com/FFmpeg/FFmpeg. Each build names the exact upstream commit it was made from inside its own version string (for example N-124464-gb2867481d9, where the trailing hash is the FFmpeg commit). The download address above tracks a rolling release, so run ffmpeg -version against your own copy to read the commit that copy was built from.
  • Build scripts, dependency versions, and the exact configure invocation used to produce the binaries: github.com/BtbN/FFmpeg-Builds.
  • The full text of the GNU General Public License v3 is installed with SiftVid, under resources/notices/licenses/, and is also published at gnu.org.

FFmpeg runs as a separate process. SiftVid invokes FFmpeg the way a person would from a terminal: it starts ffmpeg.exe or ffprobe.exe as its own operating system process and passes command-line arguments to it. SiftVid does not link against libavcodec, libavformat, or any other FFmpeg library, does not embed FFmpeg source code, and ships no FFmpeg headers or import libraries. On the mere-aggregation reading, a program that shells out to a GPL command-line tool at arm's length is not a derivative work of that tool.

No endorsement. FFmpeg is a trademark of Fabrice Bellard, originator of the FFmpeg project. Neither the FFmpeg project nor BtbN endorses, sponsors, or is affiliated with SiftVid or SiftVid. SiftVid is not part of the FFmpeg project and is not the FFmpeg project's work.

04libvips

SiftVid's image handling goes through sharp, which wraps the libvips image-processing library. Unlike FFmpeg, this one does ship inside the SiftVid installer. The platform package declares its license as Apache-2.0 AND LGPL-3.0-or-later; the LGPL half is libvips.

SiftVid uses libvips only through sharp's prebuilt native add-on, which loads it as an unmodified, dynamically linked shared library kept outside the application archive so it remains replaceable. The full text of the GNU Lesser General Public License v3 is installed with SiftVid under resources/notices/licenses/ and is published at gnu.org.

05AI model weights

Scene recognition in SiftVid is done by one model, not two: Xenova/clip-vit-base-patch32, an ONNX conversion by the transformers.js project of OpenAI's openai/clip-vit-base-patch32 (CLIP ViT-B/32). It arrives as two ONNX files — an image tower and a text tower — which are the two halves of that single model, not two separate models. The weights are used for inference only, entirely on your machine, and are never uploaded anywhere.

We do not ship these weights. Like FFmpeg, they are downloaded by your own machine from Hugging Face the first time you use auto-classification, after SiftVid tells you the size and you agree. That is a deliberate choice, for the reason in the next paragraph, and it keeps roughly 760 MB out of the installer as a side benefit.

A second model is named in the source code but does not ship. Xenova/owlvit-base-patch32 (OWL-ViT, originally published by Google Research under Apache-2.0) is reserved for a future object-spotting feature. No OWL-ViT weights are bundled in the installer or downloaded by the app, so nothing on your machine runs it and no attribution is owed for it today. If that feature ships, this page will name it before it does.

How the model affects you, legally. In our view it does not restrict you at all. The model is a fixed file of numbers that SiftVid reads in order to score frames; running it is not publication, and it produces tags in your own local database rather than a work distributed to anyone. We claim no rights in the tags, categories, or classifications the model produces on your machine — they are yours, exactly like the files they describe. Nothing about the model imposes an obligation on you to publish anything, attribute anything, or open-source anything you make with SiftVid.

Model license: undeclared upstream

Checked against the Hugging Face model API on 6 August 2026: neither the Xenova conversion nor the upstream OpenAI repository declares a license field in its model-card metadata. We will not state a license that the distributing repository does not state. OpenAI's separate CLIP reference implementation on GitHub does carry an MIT License (Copyright 2021 OpenAI), and CLIP weights are widely redistributed on the understanding that this covers them, but that understanding is not written on either model card, so we do not adopt it as fact here.

This is why SiftVid downloads the weights rather than shipping them. Redistributing a file inside a paid installer, on the strength of a licence nobody has actually granted, is not a position we are willing to take on your behalf. Fetching them from the publisher removes the question instead of guessing at the answer.

One caveat we would rather state than bury. OpenAI's CLIP model card carries a research-intent note saying that “any deployed use case of the model — whether commercial or not — is currently out of scope.” That sentence sits in the card's intended-use discussion, not in a license, and it is guidance about what the authors validated the model for rather than a grant condition binding on downstream users. We read it as a caution about accuracy and bias, not a prohibition, which is the same reading the wider ecosystem has applied to CLIP for years. It is nonetheless a fair thing for a buyer to know, and it is one reason SiftVid always presents AI classifications as suggestions you confirm rather than as verdicts.

The inference runtime around the weights is separately licensed and is listed in Section 07: @huggingface/transformers (Apache-2.0) and onnxruntime-node (MIT). Those licenses cover the runtime only, never the weights.

06What this means for you in practice

Sections 03 to 05 are written for someone checking our compliance. This section is written for someone who just bought the app and wants to know whether any of it lands on them.

Short answer: none of it does. Every obligation on this page is ours, not yours. You are an end user running software on your own computer, and none of the licenses involved place a condition on private use.

  • You owe no attribution. Nothing you export, render, or compile with SiftVid needs to credit FFmpeg, libvips, CLIP, SiftVid, or us. Exports are your files.
  • Nothing you make becomes open source. The GPL and LGPL reach the components they cover. Passing your video through a GPL-licensed encoder does not place any license on the video, and copyleft does not travel into the output of a tool.
  • The AI tags are yours. Categories, tags, scene markers, and cut lists produced on your machine belong to you. We never receive them and claim nothing in them.
  • You may keep or replace the components. The GPL and LGPL both entitle you to the source of the covered components, and to substitute your own build. The links in Sections 03 and 04 are there so you can actually do that; the components are installed as ordinary replaceable files rather than locked inside the app.
  • What you may not do is redistribute SiftVid. The one restriction that does apply to you comes from our own license, not from anyone else's: SiftVid is proprietary, and your purchase covers your use of it, not its onward distribution.

If you believe any position on this page is wrong, we would genuinely like to hear it. Email reply@siftvid.com.

07npm dependencies

Generated from the dependency tree the shipping build was compiled from. Each entry shows the license the package declares for itself; compound expressions are reproduced as written rather than reduced to one license. The list intentionally covers the whole tree rather than runtime packages only, because SiftVid's bundler compiles several development-declared libraries directly into the shipped application. Attributing a package that did not ship is harmless; omitting one that did is not.

714 packages.