Add LTX-2 marking pack (PixelSeal + AudioSeal)

718f58d verified 7 days ago

6.33 kB

---
license: mit
library_name: pytorch
tags:
  - watermarking
  - video
  - audio
  - videoseal
  - pixelseal
  - audioseal
  - ltx-2
---
# LTX-2 Marking Pack (VideoSeal / AudioSeal)

Weights and configs for the opt-in, post-hoc invisible marking used by [LTX-2](https://github.com/Lightricks/LTX-2): **PixelSeal** (VideoSeal family) for video frames and **AudioSeal** for stereo PCM.

This repository contains **weights only**. The code that loads, embeds, and detects lives in the public LTX-2 repository under [`packages/ltx-pipelines/src/ltx_pipelines/marking`](https://github.com/Lightricks/LTX-2/tree/main/packages/ltx-pipelines/src/ltx_pipelines/marking). Full documentation is in [`packages/ltx-pipelines/docs/marking.md`](https://github.com/Lightricks/LTX-2/blob/main/packages/ltx-pipelines/docs/marking.md).

> The marks are **not** cryptographic signatures. They are a robust, invisible signal that a clip was produced by an LTX-2 pipeline with marking enabled. Absence of a detected mark does not prove a clip was not generated.

## Files

The LTX-2 loader (`create_markers`) requires all five model files to be present in one directory.

| File | Description | SHA-256 |
|:--|:--|:--|
| `pixelseal.pth` | PixelSeal embedder + extractor weights (model state only) | `97cff15f75515259b49a2cffc8fcd600c287cafe3f29c1d5c6ef9c81b492c088` |
| `pixelseal.yaml` | PixelSeal model card (architecture, 256-bit payload, 256 px processing size) | `96c4a448f7d81810780e5bde5a51b039db544b7ec2baf65a605302c1b17362c9` |
| `videoseal_attenuation.yaml` | JND attenuation channel map referenced by the card | `e0d71354f8945cf600bd4ebd5e2641a9c879e017548dd6cc8a6de10239d7e821` |
| `audioseal_generator.pth` | AudioSeal 16-bit generator (embed) | `7a845b5fbe9364a63a3909d8ab3fe064d13a76ae4c2e983573e08c69b7b51748` |
| `audioseal_detector.pth` | AudioSeal 16-bit detector | `8a78e8a83584113523e161fc599fcab10fd0e94c04d2eb9d2fa1e9ec91ab69d9` |
| `LICENSE-videoseal`, `LICENSE-audioseal` | Upstream MIT licenses | |

`.pth` files are PyTorch pickles. Verify the checksums after download and load them only from this repository:

```bash
sha256sum pixelseal.pth pixelseal.yaml videoseal_attenuation.yaml audioseal_generator.pth audioseal_detector.pth
```

## Usage

### Install

Clone [LTX-2](https://github.com/Lightricks/LTX-2) and install the marking extras:

```bash
uv sync --extra videoseal --extra audioseal
# or
pip install 'ltx-pipelines[videoseal,audioseal]'
```

### Download the pack

The LTX-2 library never downloads weights itself. Fetch this repository to a local directory:

```bash
hf download Lightricks/LTX-2-Marking --local-dir path/to/marking
```

### Mark during generation

Video-generation and editing CLIs, HDR IC-LoRA, and text-to-audio accept `--marker-root-path`. Omit the flag to leave pixels and PCM unchanged.

```bash
python -m ltx_pipelines.distilled \
    ... \
    --marker-root-path path/to/marking \
    --output-path out.mp4
```

Pipeline-specific behavior:

- **HDR:** video marking is skipped; audio is still marked.
- **Audio-to-video:** only the video is marked, because the source audio is passed through.
- **Text-to-audio:** only the audio marker is used.

### Detect

```python
import torch
from ltx_pipelines.marking import create_markers, detect_audio_mark, detect_video_mark
video_marker, audio_marker = create_markers("path/to/marking", torch.device("cuda"))
print(detect_video_mark(video_marker, "out.mp4"))
print(detect_audio_mark(audio_marker, "out.mp4"))
```

Each call returns `MarkDetectResult(detected, bit_accuracy, p_value, payload)`. `bit_accuracy` and `p_value` come from a binomial test of the decoded bits against the payload, which is fixed in the LTX-2 code (not in these weights). `detected` is each marker's own decision:

- **Video:** every 8th decoded frame is resized to 256×256 and run through the extractor; bit logits are averaged over the sampled frames. `detected` is `log10(p_value) < -6`.
- **Audio:** the full track is scanned, with each stereo channel processed as its own mono signal. `detected` means at least half of the samples have a watermark probability above 0.5.

Detection runs on CPU or GPU. A GPU is much faster: on one B200, video detect takes about 15–30 ms per 5–10 s clip at up to 4K, versus 0.6–3 s on CPU. Both figures exclude file decoding.

## Scope and limitations

- Video marking targets SDR H.264 output; HDR output is not video-marked.
- AudioSeal runs at the native 24/48 kHz rate and does not resample.
- These are Meta's general-purpose models. Anyone holding the weights and the LTX-2 code can run detection, and a determined adversary may be able to weaken or remove the mark.

## Provenance

- `pixelseal.pth` is Meta's public PixelSeal checkpoint ([facebookresearch/videoseal](https://github.com/facebookresearch/videoseal)), repacked to the `model` state dict only. The optimizer, discriminator, EMA, and scheduler state were dropped; the model tensors are unchanged.
- `pixelseal.yaml` is the PixelSeal card from `videoseal` commit [`870ca7f`](https://github.com/facebookresearch/videoseal/tree/870ca7fb33578b90f14c602016b6c2788096226e). `videoseal_attenuation.yaml` is the `jnd_1_1` entry of that repo's `configs/attenuation.yaml`, which the pip package does not ship.
- `audioseal_generator.pth` and `audioseal_detector.pth` are the unmodified 16-bit base models from [facebook/audioseal](https://huggingface.co/facebook/audioseal).

## License

All weights and configs are released by Meta Platforms, Inc. under the MIT License; see `LICENSE-videoseal` and `LICENSE-audioseal`. The LTX-2 code is licensed separately; see the [LTX-2 repository](https://github.com/Lightricks/LTX-2).

## Citation

```bibtex
@article{fernandez2024video,
  title   = {Video Seal: Open and Efficient Video Watermarking},
  author  = {Fernandez, Pierre and Elsahar, Hady and Yalniz, I. Zeki and Mourachko, Alexandre},
  journal = {arXiv preprint arXiv:2412.09492},
  year    = {2024}
}
@inproceedings{sanroman2024proactive,
  title     = {Proactive Detection of Voice Cloning with Localized Watermarking},
  author    = {San Roman, Robin and Fernandez, Pierre and Elsahar, Hady and D{\'e}fossez, Alexandre and Furon, Teddy and Tran, Tuan},
  booktitle = {International Conference on Machine Learning (ICML)},
  year      = {2024}
}
```

	---
	license: mit
	library_name: pytorch
	tags:
	- watermarking
	- video
	- audio
	- videoseal
	- pixelseal
	- audioseal
	- ltx-2
	---

	# LTX-2 Marking Pack (VideoSeal / AudioSeal)

	Weights and configs for the opt-in, post-hoc invisible marking used by [LTX-2](https://github.com/Lightricks/LTX-2): PixelSeal (VideoSeal family) for video frames and AudioSeal for stereo PCM.

	This repository contains weights only. The code that loads, embeds, and detects lives in the public LTX-2 repository under [`packages/ltx-pipelines/src/ltx_pipelines/marking`](https://github.com/Lightricks/LTX-2/tree/main/packages/ltx-pipelines/src/ltx_pipelines/marking). Full documentation is in [`packages/ltx-pipelines/docs/marking.md`](https://github.com/Lightricks/LTX-2/blob/main/packages/ltx-pipelines/docs/marking.md).

	> The marks are not cryptographic signatures. They are a robust, invisible signal that a clip was produced by an LTX-2 pipeline with marking enabled. Absence of a detected mark does not prove a clip was not generated.

	## Files

	The LTX-2 loader (`create_markers`) requires all five model files to be present in one directory.

	\| File \| Description \| SHA-256 \|
	\|:--\|:--\|:--\|
	\| `pixelseal.pth` \| PixelSeal embedder + extractor weights (model state only) \| `97cff15f75515259b49a2cffc8fcd600c287cafe3f29c1d5c6ef9c81b492c088` \|
	\| `pixelseal.yaml` \| PixelSeal model card (architecture, 256-bit payload, 256 px processing size) \| `96c4a448f7d81810780e5bde5a51b039db544b7ec2baf65a605302c1b17362c9` \|
	\| `videoseal_attenuation.yaml` \| JND attenuation channel map referenced by the card \| `e0d71354f8945cf600bd4ebd5e2641a9c879e017548dd6cc8a6de10239d7e821` \|
	\| `audioseal_generator.pth` \| AudioSeal 16-bit generator (embed) \| `7a845b5fbe9364a63a3909d8ab3fe064d13a76ae4c2e983573e08c69b7b51748` \|
	\| `audioseal_detector.pth` \| AudioSeal 16-bit detector \| `8a78e8a83584113523e161fc599fcab10fd0e94c04d2eb9d2fa1e9ec91ab69d9` \|
	\| `LICENSE-videoseal`, `LICENSE-audioseal` \| Upstream MIT licenses \| \|

	`.pth` files are PyTorch pickles. Verify the checksums after download and load them only from this repository:

	```bash
	sha256sum pixelseal.pth pixelseal.yaml videoseal_attenuation.yaml audioseal_generator.pth audioseal_detector.pth
	```

	## Usage

	### Install

	Clone [LTX-2](https://github.com/Lightricks/LTX-2) and install the marking extras:

	```bash
	uv sync --extra videoseal --extra audioseal
	# or
	pip install 'ltx-pipelines[videoseal,audioseal]'
	```

	### Download the pack

	The LTX-2 library never downloads weights itself. Fetch this repository to a local directory:

	```bash
	hf download Lightricks/LTX-2-Marking --local-dir path/to/marking
	```

	### Mark during generation

	Video-generation and editing CLIs, HDR IC-LoRA, and text-to-audio accept `--marker-root-path`. Omit the flag to leave pixels and PCM unchanged.

	```bash
	python -m ltx_pipelines.distilled \
	... \
	--marker-root-path path/to/marking \
	--output-path out.mp4
	```

	Pipeline-specific behavior:

	- HDR: video marking is skipped; audio is still marked.
	- Audio-to-video: only the video is marked, because the source audio is passed through.
	- Text-to-audio: only the audio marker is used.

	### Detect

	```python
	import torch
	from ltx_pipelines.marking import create_markers, detect_audio_mark, detect_video_mark

	video_marker, audio_marker = create_markers("path/to/marking", torch.device("cuda"))
	print(detect_video_mark(video_marker, "out.mp4"))
	print(detect_audio_mark(audio_marker, "out.mp4"))
	```

	Each call returns `MarkDetectResult(detected, bit_accuracy, p_value, payload)`. `bit_accuracy` and `p_value` come from a binomial test of the decoded bits against the payload, which is fixed in the LTX-2 code (not in these weights). `detected` is each marker's own decision:

	- Video: every 8th decoded frame is resized to 256×256 and run through the extractor; bit logits are averaged over the sampled frames. `detected` is `log10(p_value) < -6`.
	- Audio: the full track is scanned, with each stereo channel processed as its own mono signal. `detected` means at least half of the samples have a watermark probability above 0.5.

	Detection runs on CPU or GPU. A GPU is much faster: on one B200, video detect takes about 15–30 ms per 5–10 s clip at up to 4K, versus 0.6–3 s on CPU. Both figures exclude file decoding.

	## Scope and limitations

	- Video marking targets SDR H.264 output; HDR output is not video-marked.
	- AudioSeal runs at the native 24/48 kHz rate and does not resample.
	- These are Meta's general-purpose models. Anyone holding the weights and the LTX-2 code can run detection, and a determined adversary may be able to weaken or remove the mark.

	## Provenance

	- `pixelseal.pth` is Meta's public PixelSeal checkpoint ([facebookresearch/videoseal](https://github.com/facebookresearch/videoseal)), repacked to the `model` state dict only. The optimizer, discriminator, EMA, and scheduler state were dropped; the model tensors are unchanged.
	- `pixelseal.yaml` is the PixelSeal card from `videoseal` commit [`870ca7f`](https://github.com/facebookresearch/videoseal/tree/870ca7fb33578b90f14c602016b6c2788096226e). `videoseal_attenuation.yaml` is the `jnd_1_1` entry of that repo's `configs/attenuation.yaml`, which the pip package does not ship.
	- `audioseal_generator.pth` and `audioseal_detector.pth` are the unmodified 16-bit base models from [facebook/audioseal](https://huggingface.co/facebook/audioseal).

	## License

	All weights and configs are released by Meta Platforms, Inc. under the MIT License; see `LICENSE-videoseal` and `LICENSE-audioseal`. The LTX-2 code is licensed separately; see the [LTX-2 repository](https://github.com/Lightricks/LTX-2).

	## Citation

	```bibtex
	@article{fernandez2024video,
	title = {Video Seal: Open and Efficient Video Watermarking},
	author = {Fernandez, Pierre and Elsahar, Hady and Yalniz, I. Zeki and Mourachko, Alexandre},
	journal = {arXiv preprint arXiv:2412.09492},
	year = {2024}
	}

	@inproceedings{sanroman2024proactive,
	title = {Proactive Detection of Voice Cloning with Localized Watermarking},
	author = {San Roman, Robin and Fernandez, Pierre and Elsahar, Hady and D{\'e}fossez, Alexandre and Furon, Teddy and Tran, Tuan},
	booktitle = {International Conference on Machine Learning (ICML)},
	year = {2024}
	}
	```