Every AI model FFmpeg GUI can use, what it is for, who made it, under which
licence, where its files come from, what the app checks them against, and
what is known of the data it was trained on. scripts/check-model-pins.py
checks before every release (release.yml) that this record holds every hash
the app pins and every licence it names, so the record and the app cannot
drift apart.
Not legal advice. This page records what the project does and why, as carefully as it can, from the licences and the pages of each model as they were when the model was added (linked below). It is not a legal opinion. If what you make with the app matters legally (a commercial product, a client's work), read the licences yourself, and ask a lawyer where it counts.
- None is in the app. The APK holds no model weights (only Parakeet's token list, a small text file: see below). FFmpeg GUI has no internet permission at all: it cannot download anything.
- The user's browser downloads it, from the page this record names, when the user taps the link in Settings, Smart features.
- The app imports it and checks it: every file against the SHA-256 pinned in the app's code (listed below), before it is ever loaded; a file that does not match is refused. "Check again" in the settings re-reads every file.
- Two kinds of file. Most are published by their makers or by a known
converter (rembg, sherpa-onnx, Xenova, Carve): the app only points to them.
Four are converted to ONNX by this project and served from its own
releases, because their makers publish them for PyTorch or TensorFlow only,
or in a form a phone cannot run: EdgeTAM, BiRefNet-lite, Bandit v2 and
Spleeter's five stems. For those the project is a distributor, and keeps
every condition of their licences (below). Two more are served from its
own releases unchanged, so that the files the app pins stay where
nobody can replace them: TrustMark's (the AI marks) and the speech model's
(Parakeet, as sherpa-onnx converted it).
Each release's page (written by
scripts/model-release-notes.py) gives the model's makers and licence (for Bandit v2, the CC BY-SA attribution and the changes), links every file it holds, the commit and the lock it was made from, its SHA-256, the paper and the sources.
| Feature | Model | Made by | Licence of the model | ONNX file by | Served from |
|---|---|---|---|---|---|
| Subtitles from speech | Parakeet TDT 0.6B v3 | NVIDIA | CC BY 4.0 | sherpa-onnx (int8) | this repository's releases (sherpa-onnx's files, unchanged) |
| Cut-out, Fast | U2-NetP | Xuebin Qin et al. | Apache 2.0 | rembg (MIT) | rembg's GitHub releases |
| Cut-out, Precise | IS-Net (general use) | Xuebin Qin et al. (DIS) | Apache 2.0 | rembg (MIT) | rembg's GitHub releases |
| Cut-out, Finest | BiRefNet-lite (general use) | Zheng Peng et al. | MIT | this project | this repository's releases |
| Cut-out correction, lift the subject | EdgeTAM | Meta | Apache 2.0 | this project | this repository's releases |
| Magic eraser, quick | MI-GAN | Picsart AI Research | MIT | its author | Hugging Face |
| Magic eraser, precise | LaMa (big-lama) | Samsung AI Center | Apache 2.0 | Carve | Hugging Face |
| Stems: dialogue, music, effects (the voice levels, the dub, the tracks) | Bandit v2 (multilingual) | Karn N. Watcharasupat, Chih-Wei Wu, Iroro Orife | CC BY-SA 4.0 (weights), Apache 2.0 (code) | this project | this repository's releases |
| The same, quicker (the default) | Bandit v2 (multilingual), int8 | Karn N. Watcharasupat, Chih-Wei Wu, Iroro Orife | CC BY-SA 4.0 (weights), Apache 2.0 (code) | this project (int8, from its float file) | this repository's releases |
| Stems: the music's instruments | Spleeter 5 stems | Deezer | MIT | this project | this repository's releases |
| AI dubbing | Kokoro-82M v1.0 | hexgrad | Apache 2.0 | sherpa-onnx (Apache 2.0) | sherpa-onnx's GitHub releases |
| AI dubbing, extra voices | Piper voices | their authors | each its own | Piper / sherpa-onnx | Hugging Face, GitHub |
| Subtitle translation | Opus-MT | University of Helsinki | CC BY 4.0 | Xenova | Hugging Face |
| AI marks (voices), before the dub's and the eraser's models | AudioSeal | Meta | MIT | this project | this repository's releases |
| AI marks (pictures), before the dub's and the eraser's models | TrustMark Q | Adobe | MIT | Adobe | this repository's releases (Adobe's files, unchanged) |
EdgeTAM ("EdgeTAM: On-Device Track Anything Model", Meta, CVPR 2025) is a Segment Anything 2 made for phones: given points on a picture marked "this is part of it" or "this is not", it finds the whole thing they point at. The app uses it to correct a cut-out with brushes, in the magic eraser's and the HDR look's magic brushes, and in the photo editor's long press that lifts a subject off the picture as a sticker. It replaced SlimSAM (see "Retired models").
-
Licence. "The EdgeTAM model checkpoints and code are licensed under Apache 2.0" (the repository's README); its
LICENSEis the Apache License 2.0, SHA-256c71d239df91726fc519c6eb72d318ec65820627232b2f796219e87dcf35d0ab4. -
Made from, all pinned in
scripts/edgetam-onnx.lock: the repository https://github.com/facebookresearch/EdgeTAM at commit7711e012a30a2402c4eaab637bdb00a521302c91, its checkpointcheckpoints/edgetam.pt(SHA-256ed2d4850b8792c239689b043c47046ec239b6e808a3d9b6ae676c803fd8780df) and its configconfigs/edgetam.yaml. -
What the conversion does (
scripts/export-edgetam.py): it loads Meta's checkpoint strictly (every weight from it; the ImageNet weights the backbone would fetch through timm are not used), cuts the model in two where SAM 2's own image predictor cuts it (the image encoder; the prompt encoder with the mask decoder) and writes both as ONNX, opset 17. No weight is trained, fine-tuned, pruned or quantised. Each run checks the files against the PyTorch model, on random inputs and on real photos, and fails rather than write files that answer differently. -
Reproducible. Constant folding is off and every tool is pinned by hash (
scripts/model-export-requirements.txt): two runs give the same bytes. The workflow that publishes them builds them twice, on two runner images, and publishes only byte-identical results. -
The files, pinned in
app/.../cutout/RefineModel.kt:File SHA-256 Bytes edgetam-image-encoder.onnx692c9365d13fbbd8fd5b32bdd0c7bd0377696bb0b20fa7d97f8a3a3c1e776f0d20,164,359 edgetam-mask-decoder.onnxdf71726f03f1eb11ab0fda4648ad024569fbbe26ddd69e2a16354432158035a216,574,577 Served as the pre-release
edgetam-onnx-1of this repository, made by.github/workflows/models-onnx.yml. -
How the Apache License 2.0 is kept (its section 4, for a Derivative Work):
- (a) a copy of the licence goes with the files:
LICENSEin the release, inside each ONNX file (its metadata, keylicense), and in the app (assets/licenses/edgetam/LICENSE.txt, readable in About, Licences); - (b) prominent notices that the files were changed:
NOTICEin the release says what the conversion changed; the same text is the ONNX files'doc_stringand theirnoticemetadata, and is in the app (assets/licenses/edgetam/NOTICE.txt); - (c) the copyright notices are kept: "Copyright (c) Meta Platforms, Inc. and affiliates", as the source files carry it, is in the NOTICE;
- (d) the upstream NOTICE file: the EdgeTAM repository has none;
- section 6, trademarks: Meta's and EdgeTAM's names are used only to say where the files come from; the NOTICE says Meta has not reviewed or endorsed them;
- sections 7 and 8: the files come with no warranty, as the licence says.
- (a) a copy of the licence goes with the files:
-
Training data. From the paper: the image encoder was trained on SA-1B ("we train on the SA-1B dataset for 2 epochs"); the video stages on SA-V, a 10% sample of SA-1B, DAVIS, MOSE and YouTube-VOS, distilling from SAM 2. SA-1B is published by Meta under its own data licence, for research; SA-V under CC BY 4.0 (the SAM 2 paper). Those terms are for whoever uses the datasets: this project uses none of them, only the checkpoint, under the licence Meta gave it.
BiRefNet ("Bilateral Reference for High-Resolution Dichotomous Image Segmentation", Zheng Peng et al., CAAI AIR 2024), the "lite" model of its author's general-use weights (Swin-T backbone): it finds the subject of a picture with finer edges than IS-Net (hair, fur, thin parts come out whole) and keeps to the main subject. It is the cut-out's "Finest" choice.
-
Licence. MIT ("Copyright (c) 2024 ZhengPeng"); its
LICENSE, SHA-25692a7089e0915fc32bc40067560b398f1e6a7a5958abd7d04eda393629a5acefb. -
Made from, all pinned in
scripts/birefnet-onnx.lock: the repository https://github.com/ZhengPeng7/BiRefNet at commitf7079632d681ecf12161fcea8905682934e59e51, and the author's weightsBiRefNet-general-bb_swin_v1_tiny-epoch_232.pthfrom the repository's release v1 (SHA-2566a1e050c6ec2697e5ed268455df544782b023acf8643ab771250979094875ab1). -
Why our own file. The author also publishes an ONNX of these weights (release v1, opset 17, SHA-256
5600024376f572a557870a5eb0afb1e5961636bef4e1e22132025467d0f03333), but it writes the decoder's deformable convolutions as gathers, which need more than 5.7 GB of memory at 1024 x 1024: no phone runs it. Ours writes them as ONNX's ownDeformConv(opset 19), which ONNX Runtime 1.28 (the app's) runs natively: the same convolution, about 2.1 GB at its peak. -
What the conversion does (
scripts/export-birefnet.py): it picks the lite backbone in a copy of the author'sconfig.py, loads the weights strictly, keeps the last (finest) prediction as the output and writes ONNX, opset 19. No weight is trained, fine-tuned, pruned or quantised. Each run checks the file against the PyTorch model (random inputs and real photos), and can check it against the author's own ONNX too. -
The file, pinned in
app/.../cutout/CutoutModel.kt(FINEST):File SHA-256 Bytes birefnet-lite-general.onnx8b75d639b11e984fc7d3a6c0818a2dd2f62df0b19997e4e822099e72f8d9b774180,240,681 Served as the pre-release
birefnet-lite-onnx-1of this repository, made by.github/workflows/models-onnx.yml, built twice to the same bytes as EdgeTAM's. -
How the MIT License is kept: "the above copyright notice and this permission notice shall be included in all copies or substantial portions" of it: the
LICENSEis in the release, inside the ONNX file (metadatalicense) and in the app (assets/licenses/birefnet/LICENSE.txt); aNOTICEsays what the conversion did (MIT does not ask for it; it is said all the same), also inside the file and in the app. -
Training data. The author lists, for these weights: DIS5K-TR, DIS-TEs, DUTS-TR_TE, HRSOD-TR_TE, UHRSD-TR_TE, HRS10K-TR_TE, TR-P3M-10k, TE-P3M-500-NP, TE-P3M-500-P and TR-humans (the repository's README, "model zoo"). These datasets come with terms of their own (DIS5K's, for one, asks commercial users to sign separate terms). They bind whoever uses the datasets: this project uses none of them, only the weights, under the licence their author gave them.
Bandit v2 ("Remastering Divide and Remaster: A Cinematic Audio Source Separation Dataset with Multilingual Support", Karn N. Watcharasupat, Chih-Wei Wu and Iroro Orife, IEEE IS2 2024) splits any sound into its dialogue, its music and its effects. The app uses it to set the voice apart (the voice levels, the dub keeping or lowering the original voice) and to lay each part on the video as a track of its own. It replaced Spleeter's two stems (see "Retired models").
-
Licences. The weights: Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0), as Zenodo record 12701995 gives them; its legal code as SPDX keeps it (
CC-BY-SA-4.0.txt, SPDX license list v3.27.0), SHA-256cde7883b9050a1104f4ac19a1572aafd6e5d7323b68351aaf51fbf4beba54966. The code: Apache 2.0, itsLICENSESHA-256c71d239df91726fc519c6eb72d318ec65820627232b2f796219e87dcf35d0ab4. CC BY-SA allows commercial use; it asks for the attribution, a link to the licence, a note of the changes, and that an adaptation goes out under the same licence: the ONNX file is an adaptation and goes out under CC BY-SA 4.0 too. It reaches only the model's file, never the sound a user makes with it, nor FFmpeg GUI's own code. -
Made from, all pinned in
scripts/bandit-onnx.lock: the repository https://github.com/kwatcharasupat/bandit-v2 at commitd5563d9031e95fdaa3e5a73d5020b9a0df61adb6and the multilingual checkpointcheckpoint-multi.ckpt(MD5fea2868787551b0cff36cfcf7c3622a3, SHA-256abcfccf65446752a057f4a302c941479a54b7560ebf8d7bca039d2ea98e64cfc), the MD5 as Zenodo gives it, the SHA-256 learned by the first build (scripts/pin-stems.py). -
Why our own file. The authors publish a PyTorch Lightning checkpoint only, and their model builds its masks as complex tensors, which ONNX does not have.
-
What the conversion does (
scripts/export-bandit.py): it builds the authors' model asconfigs/models/bandit-mus64.yamldoes (48 kHz), loads the checkpoint strictly, and writes the network from the spectrogram to the three complex masks with real numbers only: the same layers, in the same order, with the same weights, ONNX opset 17. The short-time Fourier transforms stay out of the file: the app does them (core/stems/StemSeparation.kt,BanditMasks.kt), as the script'sstft()andistft(), which each build checks against the authors' own model end to end. No weight is trained, fine-tuned, pruned or quantised. The model is mono: the app runs it on the middle of the two channels and lays its masks on each channel, so the stems stay stereo. -
The file, pinned in
app/.../voice/StemModels.kt(BANDIT):File SHA-256 Bytes bandit-v2-multi.onnx7836f01622021adc2fe2570f263ec365ef7677d598b51d27b015b94d869f2ff2149,293,980 Zeros until its first build:
.github/workflows/models-onnx.yml(without publishing) builds it twice, the two must be byte-identical, thenscripts/pin-stems.py bandit <the build's folder>writes the pins here, in the lock, in the app and in the bundled NOTICE; the next run publishes the pre-releasebandit-onnx-2. The first file,bandit-onnx-1(SHA-2561601a8dc45ddea3b0d3d16378c9e0d42ba7e2b6cef3bd832d5f8e93808a65fd4, 150,610,260 bytes), put the masks together band by band, with the authors' float64 frequency weights: some 23,000 nodes, 12 s to load on a computer, masks as doubles. bandit-onnx-2 is the same weights with one scatter-add per stem and float32 throughout: under 9,000 nodes, about 3 times quicker to load, some 20% quicker to run. The app retires the first (see "Retired models"). Until then the app says the model is on its way and imports nothing. -
How CC BY-SA 4.0 is kept: the attribution (the authors, the title of the work, the record's address), the licence's legal code and the note of the changes are in the release (
LICENSE.txt,NOTICE.txt), inside the ONNX file (metadatalicense,notice) and in the app (assets/licenses/bandit/, shown in About, Licences); no technical measure restricts the file. -
Training data. Divide and Remaster v3, the authors' dataset: synthetic mixtures of speech in more than 30 languages, music and effects, which the paper says it drew only from content that permits commercial use (for the music, the Free Music Archive's public-domain and CC BY, CC BY-SA, Open Audio and Free Art tracks). This project uses none of it, only the weights.
The same weights as bandit-onnx-2, with their matrix products in 8-bit
integers: the build the app runs by default. On a phone, splitting four
minutes of sound with the float file took some twenty minutes; the int8 one
loads about three times quicker and runs quicker where the processor has
int8 instructions (1.7 times on the desktop processor it was first measured
on, a 2.1 GHz Xeon with VNNI; a phone's gain is measured on the phone).
The float file stays offered, below it, for the two cases where it is worth
it: an older phone without int8 instructions (from before about 2018, where
the int8 build may be no quicker), or stems that must match the authors'
model to the last bit.
-
Licences. As the float file: an adaptation of the CC BY-SA 4.0 weights, under CC BY-SA 4.0 too, with the same attribution and the legal code next to it and inside it (
assets/licenses/bandit-int8/in the app). -
Made from, all pinned in
scripts/bandit-int8-onnx.lock: the float file of the releasebandit-onnx-2(SHA-2567836f01622021adc2fe2570f263ec365ef7677d598b51d27b015b94d869f2ff2, which the script checks against both locks), the CC BY-SA 4.0 legal code, and ONNX Runtime 1.31.1, whose quantisation writes the bytes (the script refuses another version). Nothing comes from the checkpoint again. -
What the script does (
scripts/export-bandit-int8.py): it folds the float file's constants, writes each of its 16 GRU layers (bidirectional, 256 units) out as what it computes (the input's matrix product for every frame at once, then a scan over the frames whose step is the GRU cell), which answers as the float file does to rounding (checked: at most 1e-4), then quantises every matrix product's weights to signed 8-bit integers, one scale per output channel, those inside the scans too (ONNX Runtime's dynamic quantisation; the values in between are quantised as the model runs). Layer norms, gates and the masks' sums stay in float. ONNX Runtime's GRU kernel cannot be quantised: that is why the layers are written out. -
How close it stays, checked by every build (its
check.json): the masks of random spectrograms differ from the float file's by about 0.2% on average (at most 2%), and each of the three stems of a test sound (tones, noise, chirps, clicks) stays about 40 dB from the float one (at least 18). On a mix of speech, music and effects whose parts are known, the float file found the dialogue at 21.30 dB SDR and the int8 one at 21.29. -
The file, pinned in
app/.../voice/StemModels.kt(BANDIT_INT8):File SHA-256 Bytes bandit-v2-multi-int8.onnx3a4a054b8e07c553d5f19f9f72ad52a1d5049c6355e0512f33b9e2f5ae7579a940,260,047 Zeros until its first build:
.github/workflows/models-onnx.yml(without publishing) builds it twice, the two must be byte-identical, thenscripts/pin-stems.py bandit-int8 <the build's folder>writes the pins here, in the lock, in the app and in the bundled NOTICE; the next run publishes the pre-releasebandit-int8-onnx-1. Until then the app keeps to the float file and says the int8 one is on its way.
Spleeter ("Spleeter: a fast and efficient music source separation tool with pre-trained models", Romain Hennequin et al., Deezer Research, JOSS 2020), its five-stem model: singing, piano, drums, bass and the rest. The app runs it on the music Bandit finds, to lay each instrument on the video as a track of its own.
-
Licence. MIT ("Copyright (c) 2019-present, Deezer SA"); its
LICENSE, SHA-256b15e40c921ec6fe997e0d88ddab9c26fb915bebbbb9496530d3f98f317d4325a. -
Made from, all pinned in
scripts/spleeter-onnx.lock: the repository https://github.com/deezer/spleeter at commit6f2c8a2e283914faa544631430f65b793f1b1218, Deezer's TensorFlow checkpoints5stems.tar.gzand (to check the conversion)2stems.tar.gzof release v1.4.0, and sherpa-onnx's ONNX of the two-stem model, as the lock pins them. -
Why our own file. Deezer publishes TensorFlow checkpoints only, and sherpa-onnx converted the two-stem model alone.
-
What the conversion does (
scripts/export-spleeter.py): it reads the checkpoints' weights without TensorFlow (the tensor bundle format) into the same five U-Nets written in PyTorch (ELU, softmax across the instruments, as Spleeter'sconfigs/5stems) and writes them as one ONNX file, opset 17. The same code must give, on the two-stem model, what sherpa-onnx's files give; then the five-stem file is checked against PyTorch. No weight is trained, changed or quantised. -
The file, pinned in
app/.../voice/StemModels.kt(INSTRUMENTS_FILE):File SHA-256 Bytes spleeter-5stems.onnx533f0f260e6bfd14f55027594ba443e9d5c1cd8a1507a67000fa312d5b2374cf196,647,290 Served as the pre-release
spleeter-onnx-1of this repository, made by.github/workflows/models-onnx.yml, built twice to the same bytes. -
How the MIT License is kept: the
LICENSEis in the release, inside the file (metadatalicense) and in the app (assets/licenses/spleeter/), with aNOTICEof the conversion. -
Training data. Deezer's own, as its paper says (not published). This project uses only the weights, under the licence Deezer gave them.
What the app's models make carries a mark that cannot be seen or heard (EU
AI Act, Regulation (EU) 2024/1689, Article 50(2); see docs/ARCHITECTURE.md,
"Marked as made with AI"). These four files write and find it. They are
imported before the models that make voices and pictures, the dub's (Kokoro,
the Piper voices) and the magic eraser's (MI-GAN, LaMa) (MarkingGate);
without them none of those stays on the phone (MarkingSweep). The other
models (speech, translation, cut-out and its corrections, stems) make
nothing that is marked and need no marks. The marks carry nothing but "made with AI":
their payloads are constants, the same for everyone (core/marking/Marks.kt).
That they carry no fingerprint is the design, not a certainty: see
docs/ARCHITECTURE.md, "That they carry nothing is not certain".
-
AudioSeal ("Proactive Detection of Voice Cloning with Localized Watermarking", Meta, ICML 2024), which marks voices. Licence: MIT, "We have updated our license to full MIT license (including the license for the model weights)" (its README); its
LICENSE, SHA-256da6d3703ed11cbe42bd212c725957c98da23cbff1998c05fa4b3d976d1a58e93. Made from, pinned inscripts/ai-markers.lock: the packageaudioseal==0.2.1(the network's code) and the weightsgenerator_base.pth, SHA-2567a845b5fbe9364a63a3909d8ab3fe064d13a76ae4c2e983573e08c69b7b51748, anddetector_base.pth, SHA-2568a78e8a83584113523e161fc599fcab10fd0e94c04d2eb9d2fa1e9ec91ab69d9(https://huggingface.co/facebook/audioseal). What the conversion does (scripts/export-ai-markers.py): the generator asget_watermarkwith its message as an input, the detector asforward, both written as ONNX (opset 17) for a fixed window of 5 s at 16 kHz; nothing trained, fine-tuned, pruned or quantised; checked against PyTorch, and checked to hold on speech through MP3, AAC, noise and volume and not to be found where it is not. Training data. "Our watermark generator and detector are trained on a 4.5K hours subset from the VoxPopuli (Wang et al., 2021) dataset" (the paper): speech of the European Parliament, published by Meta for research. -
TrustMark Q ("TrustMark: Robust Watermarking and Watermark Removal for Arbitrary Resolution Images", T. Bui, S. Agarwal, J. Collomosse, Adobe and the University of Surrey, ICCV 2025), which marks pictures. Licence: MIT, "The MIT license covers both the code in this repo and the model files which are downloaded upon first use" (its README); its
LICENSE, SHA-25658d47a5cdfaaceda98925d845e8207d90336c62347a2383f95962b667b1cdaf5. Adobe confirmed that the files may be redistributed unchanged, with this attribution: adobe/trustmark#48. Made from: Adobe's own ONNX files, unchanged byte for byte, renamed only:encoder_Q.onnx, SHA-25619b3d1b25836130ffd78775a8f61539f993375d1823ef0e59ba5b8dffb4f892d, anddecoder_Q.onnx, SHA-256ee3268f057c9dabef680e169302f5973d0589feea86189ed229a896cc3aa88df(https://cai-watermark.adobe.net/watermarking/trustmark-models/, the location in Adobe's own code), Adobe's reader checked with TrustMark at commitf0df43ed760bf3305f98f495f990b8c2315e7983. Training data. "We follow the same settings of [7] to train our models on 101K images from the MIRFlickR 1M dataset" (the paper). -
The files, pinned in
app/.../marking/MarkerModel.kt:File Made from SHA-256 Bytes audioseal-generator.onnxAudioSeal, converted ecb230f2d38de821f72f04eda9c40b82f102e058263848f892be4f6e189dd0e758,960,229 audioseal-detector.onnxAudioSeal, converted a389f9a5e3fa67a6a3a13480c3799875a5f3251d2d542fada6bd93b5c2bb53fa34,754,353 trustmark-encoder-q.onnxAdobe's encoder_Q.onnx19b3d1b25836130ffd78775a8f61539f993375d1823ef0e59ba5b8dffb4f892d17,312,208 trustmark-decoder-q.onnxAdobe's decoder_Q.onnxee3268f057c9dabef680e169302f5973d0589feea86189ed229a896cc3aa88df47,401,222 A value of zeros is not pinned yet:
.github/workflows/ai-markers.ymllearns it on its first run (building and checking everything, publishing nothing),scripts/pin-ai-markers.pycopies what it learned here, into the lock and into the app; the next run publishes the pre-releaseai-markers-1. Until then nothing can be imported (no file matches a pin of zeros), andscripts/check-model-pins.pyrefuses an app release. -
How the MIT licences are kept: each
LICENSEwith its copyright notice goes with the files (LICENSE-audioseal.txt,LICENSE-trustmark.txtin the release; inside the AudioSeal ONNX files, metadata keylicense) and is in the app (assets/licenses/ai-markers/, readable in About, Licences), with the NOTICE of what was converted and what is unchanged.
By NVIDIA, CC BY 4.0 (https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3), as
the sherpa-onnx project exported it to ONNX and quantised it to int8
(https://github.com/k2-fsa/sherpa-onnx; published at
https://huggingface.co/csukuangfj/sherpa-onnx-nemo-parakeet-tdt-0.6b-v3-int8
and in sherpa-onnx's asr-models release). CC BY asks for credit: the app
gives it in About, Licences, and in Smart features. Its token list ships with
the app (assets/licenses/parakeet/ATTRIBUTION.txt, with the licence).
Why this repository serves it. The places upstream serve whatever was
uploaded last under the same name. In November 2026 what was downloaded from
them stopped passing the app's check against the pins it has had since
2026-10-18 (commit 1fbba8d), the archive and the loose files alike, and why
was not established. The app refused them, as it should, and its pins were
not moved to whatever is served now: a file that changes without a word is
exactly what they are there to stop. Instead the files the app was pinned
to, checked against those pins, are served from this repository's release
parakeet-onnx-1, byte for byte, where they cannot change under the app.
scripts/parakeet-onnx.lockpins the three files (name, SHA-256, size), the release tag and the licence;speech/SpeechModel.ktpins the same, andscripts/check-model-pins.pychecks that they agree.- The release holds the three files with
LICENSE.txt(the CC BY 4.0 legal code, the copy the app bundles),NOTICE.txt(the attribution, where the files come from, and that nothing was changed) andSHA256SUMS.txt, all three written byscripts/check-model-pins.py --parakeet-side-files DIR;--release-dir parakeet=DIRchecks a folder before it is uploaded. - CC BY 4.0, section 3(a): the release and the app credit NVIDIA, keep its notices, link the licence, say that the files are unchanged and that no endorsement is implied.
The files: encoder.int8.onnx acfc2b4456377e15d04f0243af540b7fe7c992f8d898d751cf134c3a55fd2247
(652,184,281 bytes), decoder.int8.onnx
179e50c43d1a9de79c8a24149a2f9bac6eb5981823f2a2ed88d655b24248db4e (11,845,275
bytes), joiner.int8.onnx
3164c13fc2821009440d20fcb5fdc78bff28b4db2f8d0f0b329101719c0948b3 (6,355,277
bytes), and in the app tokens.txt
d58544679ea4bc6ac563d1f545eb7d474bd6cfa467f0a6e2c1dc1c7d37e3c35d (kept byte
for byte by .gitattributes).
No longer offered. The SmoothQuant build (by pantinor, on Hugging Face,
CC BY 4.0: encoder.int8.onnx
1db13b37849829eecce166bf2e1c7d4d334ceed9e3b9b95c515188dd870301ae,
decoder.int8.onnx
0018457edb28af7715cfd86ae47a6d67775492d7d32de58a41e7aa09305121e1) and the
import of sherpa-onnx's archive
(5793d0fd397c5778d2cf2126994d58e9d56b1be7c04d13c7a15bb1b4eafb16bf), with
the speed test that listened to the archive's recordings, are gone: a
SmoothQuant model still on a phone is deleted at the next start
(RetiredModels), and so are the recordings.
The app only links to these and checks them; it does not host or change them.
By Xuebin Qin et al. (https://github.com/xuebinqin/DIS, https://github.com/xuebinqin/U-2-Net), as converted to ONNX by rembg (Daniel Gatis, MIT, https://github.com/danielgatis/rembg) and served from rembg's GitHub releases. Both repositories are under Apache 2.0; DIS's page does not name a separate licence for its weights, which are taken here, as rembg takes them, as covered by the repository's.
isnet-general-use.onnx60920e99c45464f2ba57bee2ad08c919a52bbf852739e96947fbb4358c0d964a(178,648,008 bytes).u2netp.onnx309c8469258dda742793dce0ebea8e6dd393174f89934733ecc8b14c76f4ddd8(4,574,861 bytes).- Trained on DIS5K (IS-Net) and DUTS-TR (U2-Net), datasets with their own terms (see BiRefNet above for DIS5K's).
- MI-GAN, by Picsart AI Research, code and weights MIT
(https://github.com/Picsart-AI-Research/MI-GAN); its ONNX pipeline is its
author's, on Hugging Face (andraniksargsyan/migan):
migan_pipeline_v2.onnx6f1f3530a1a2324b19752018ce756088b07973cda8d7d890034ace5c8a48c40b. - LaMa ("big-lama"), by Samsung AI Center, Apache 2.0 (https://github.com/advimman/lama), as converted to ONNX by Carve (huggingface.co/Carve/LaMa-ONNX). Not published with a hash: the app takes its SHA-256 when it is imported and shows it, and checks it against that afterwards.
- Both learnt from Places2, a set of photos published for research.
By hexgrad, Apache 2.0 (https://huggingface.co/hexgrad/Kokoro-82M), in the
multi-language build of sherpa-onnx (Apache 2.0), with eSpeak NG's data
(GPL 3.0 or later) and jieba's dictionaries (MIT): the archive
kokoro-multi-lang-v1_0.tar.bz2
c5f7e2d2caf082bc1d20fb70334a61d99d20b484500aad32e7cf84c128ea3298, holding
model.onnx b40f62b166ac8164b0627ef48a0b358eda0985e272fb03ef5252e7206305da11,
voices.bin 1c5a5b983d3d50d8586d437a51f3faa2da7919ce76a013c081e65671a3447c29,
tokens.txt 6ebb6bb288f20f3ae8d004d3c2ca27697da27c037d75e81a60e2a6a663f95425,
lexicon-us-en.txt 7daaab53a181be9885b853a8582bf1838186317e5dadacbcef9c426d6fa0da14,
lexicon-gb-en.txt c4cbb37316f62210dff52718a7afcaae24f50c032cc75ab47ae67b831d1049e7,
lexicon-zh.txt 902bc2d20ac7c449c5ecbbbb23c89d12119c03ed5b8ed27f114f0260c9e35229
and eSpeak NG's phontab 886f3fa402cb0ba73d483aa8ad000af47a6b7cc06293c75a97913fba68a530f6.
Licences in the app: assets/licenses/kokoro/, assets/licenses/sherpa-onnx/espeak-ng/.
Each voice is under the licence its model card states; the app reads the card
and, for a voice not free for every use (not for commercial use, research
only, or not stated), says so before every dub with it, and keeps the
responsibility with whoever publishes the video. Piper itself is MIT
(assets/licenses/piper/LICENSE.txt). Their hashes are taken at import.
By the University of Helsinki (Jorg Tiedemann et al.), CC BY 4.0 (https://github.com/Helsinki-NLP/Opus-MT), as converted to ONNX by Xenova on Hugging Face. Any language pair can be added, so their hashes are taken at import and checked again on request. CC BY asks for credit: About, Licences.
Models the app used before. When a newer version starts and finds one still
on the phone, it deletes it (it would never be loaded again) and says so once,
with what took its place and how much space came back
(app/.../cutout/RetiredModels.kt, RetiredModelsDialog.kt).
| Model | Licence | Was used for | Folder deleted | Replaced by |
|---|---|---|---|---|
| SlimSAM-77 (Zigeng Chen et al., ONNX by Xenova) | Apache 2.0 | cut-out correction | files/refine/slimsam-77 |
EdgeTAM |
Bandit v2's first file, bandit-onnx-1 (see above) |
CC BY-SA 4.0 | stems | files/stems/bandit-v2 |
Bandit v2, bandit-onnx-2 |
Spleeter 2stems (Deezer, MIT; int8 ONNX by sherpa-onnx, Apache 2.0: sherpa-onnx-spleeter-2stems-int8.tar.bz2 c5dc8edfa08cc9a6d82670fcca15cfe671e446385e1a28c9da8c1816d238ea17, vocals.int8.onnx 6e9721864c0c1a6d86eaae92de71637b2daeb3df769c9c536053b46be8dd10aa, accompaniment.int8.onnx 53a96a0d1520b45b8f3cc53262e3725defb686fc3a17490932f050721f4583f1) |
MIT | voice separation | files/voice/spleeter-2stems-int8 |
Bandit v2, and Spleeter 5 stems for the instruments |
The voices Spleeter's two stems set apart stay (they are results, in
files/voice-stems): projects keep using them; splitting such a file into
tracks makes it again with Bandit.
- RMBG-1.4 and RMBG-2.0 (BRIA AI): very good cut-outs, but their weights are for non-commercial use only (RMBG-2.0 is BiRefNet trained on BRIA's own data). FFmpeg GUI's users may well work commercially: left out.
- ViTMatte and other matting models trained on Adobe's Composition-1k: that dataset's terms are for research only, and the weights' own licence is unclear. Left out until both are clear. BiRefNet's own matting weights are MIT, but work at 2048 x 2048, beyond a phone's memory.
- SAM 2 / SAM 2.1 (Meta, Apache 2.0): far heavier than EdgeTAM for the same job on a phone.
- The author's ONNX of BiRefNet-lite: the same licence, but more than 5.7 GB of memory (see above).
- Read its licence and its model card, and those of its ONNX file if someone else converted it. Note what it was trained on. A licence that forbids commercial use, or none at all, keeps a model out unless the app warns before every use (as for Piper voices).
- Pin every file by SHA-256 (and size where it is known) in its Kotlin
object, and its licence in the object (
LICENCE,licence). - Record it here: what for, who, licence, source, every hash, training data.
- Credit it where its licence asks: the README's Credits, About, Licences
(
settings/AboutScreen.kt),assets/licenses/README.txt("Not part of the APK"), and its licence text inassets/licenses/when the app shows it. - If the project converts it: a lock (
scripts/*-onnx.lock), an export script that writesLICENSE,NOTICEandPROVENANCEnext to the files and the licence and notice inside them, the workflow (models-onnx.yml), and the bundled notice (assets/licenses/<model>/NOTICE.txt, the same text). - If it replaces a model, add the old one to
RetiredModels.ALL(never take one away) and to "Retired models" above. - Run
scripts/check-model-pins.py: it fails until 2 to 6 agree.
python3.13 -m venv venv
venv/bin/pip install --require-hashes --no-deps -r scripts/model-export-requirements.txt
git clone https://github.com/facebookresearch/EdgeTAM && git -C EdgeTAM checkout 7711e012a30a2402c4eaab637bdb00a521302c91
venv/bin/python -I scripts/export-edgetam.py --edgetam EdgeTAM --out out-edgetam --pictures EdgeTAM/notebooks/images/truck.jpg
git clone https://github.com/ZhengPeng7/BiRefNet && git -C BiRefNet checkout f7079632d681ecf12161fcea8905682934e59e51
curl -LO https://github.com/ZhengPeng7/BiRefNet/releases/download/v1/BiRefNet-general-bb_swin_v1_tiny-epoch_232.pth
venv/bin/python -I scripts/export-birefnet.py --birefnet BiRefNet --weights BiRefNet-general-bb_swin_v1_tiny-epoch_232.pth --out out-birefnet
python3 scripts/check-model-pins.py --release-dir edgetam=out-edgetam --release-dir birefnet=out-birefnet
# The stems (the upstream files as the locks and models-onnx.yml say):
git clone https://github.com/kwatcharasupat/bandit-v2 && git -C bandit-v2 checkout d5563d9031e95fdaa3e5a73d5020b9a0df61adb6
venv/bin/python -I scripts/export-bandit.py --bandit bandit-v2 --checkpoint checkpoint-multi.ckpt --licence CC-BY-SA-4.0.txt --out out-bandit
venv/bin/python -I scripts/export-spleeter.py --checkpoint 5stems --two-stems 2stems --sherpa sherpa-onnx-spleeter-2stems --licence spleeter/LICENSE --out out-spleeter
python3 scripts/check-model-pins.py --release-dir bandit=out-bandit --release-dir spleeter=out-spleeter
# The AI marks (needs espeak-ng and ffmpeg; the upstream files as the lock and ai-markers.yml say):
venv/bin/python -I scripts/export-ai-markers.py --generator generator_base.pth --detector detector_base.pth \
--encoder encoder_Q.onnx --decoder decoder_Q.onnx --audioseal-licence LICENSE-audioseal --trustmark trustmark \
--pictures EdgeTAM/notebooks/images/truck.jpg --out out-markers
python3 scripts/check-model-pins.py --release-dir markers=out-markers
The last line says whether your files are the very ones the app pins.