{"id":14094,"date":"2026-11-17T11:02:53","date_gmt":"2026-11-17T02:02:53","guid":{"rendered":"https:\/\/localmodelwatch.tsuchitsuchi.com\/?p=14094"},"modified":"2026-11-17T11:02:53","modified_gmt":"2026-11-17T02:02:53","slug":"lightricks-ltx-2-marking-invisible-watermarking","status":"publish","type":"post","link":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/lightricks-ltx-2-marking-invisible-watermarking\/","title":{"rendered":"LTX-2-Marking Video and Audio Watermarking Model: File List"},"content":{"rendered":"<p><!-- lmw:facts --><\/p>\n<h2>At a Glance<\/h2>\n<div class=\"lmw-table-scroll\" tabindex=\"0\" style=\"overflow-x:auto;-webkit-overflow-scrolling:touch;max-width:100%;\">\n<table style=\"width:max-content;min-width:100%;border-collapse:collapse;\">\n<thead>\n<tr>\n<th>Item<\/th>\n<th>Value<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>Repository<\/td>\n<td><a href=\"https:\/\/huggingface.co\/Lightricks\/LTX-2-Marking\">Lightricks\/LTX-2-Marking<\/a><\/td>\n<\/tr>\n<tr>\n<td>Publisher guide<\/td>\n<td><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/publisher-lightricks-en\/\">Lightricks (LTX): models and licenses<\/a><\/td>\n<\/tr>\n<tr>\n<td>Published<\/td>\n<td>2026-10-28<\/td>\n<\/tr>\n<tr>\n<td>License<\/td>\n<td>mit<\/td>\n<\/tr>\n<tr>\n<td>Formats<\/td>\n<td>PyTorch (pth)<\/td>\n<\/tr>\n<tr>\n<td>Paper<\/td>\n<td><a href=\"https:\/\/arxiv.org\/abs\/2412.09492\">arXiv:2412.09492<\/a><\/td>\n<\/tr>\n<tr>\n<td>Source type<\/td>\n<td>Primary source (the publisher itself)<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p><em>Values determined by this site&#8217;s code when the information was collected. Dates are JST.<\/em><\/p>\n<p><!-- \/lmw:facts --><\/p>\n<h2>Overview<\/h2>\n<p>Lightricks has released &#8220;Lightricks\/LTX-2-Marking&#8221;, a collection of model weights and configuration files for the invisible electronic watermarking (invisible marking) used in their video generation pipeline &#8220;LTX-2&#8221;.<\/p>\n<p>This pack consists of &#8220;PixelSeal&#8221; (part of the VideoSeal family), a watermarking technology for video frames, and &#8220;AudioSeal&#8221; for stereo PCM audio. Signals that are imperceptible to human eyes and ears are post-effectively embedded into the generated video and audio signals, and these signals can later be detected to determine whether the target media was generated by the LTX-2 pipeline with watermarking enabled. The publisher explains that this feature is not a cryptographic signature, and even if a watermark is not detected, it does not serve as complete proof that the content &#8220;was not created by a generative model&#8221;.<\/p>\n<h2>Specifications<\/h2>\n<p>The main specifications and configuration files listed in the published documentation are as follows:<\/p>\n<ul>\n<li><strong>Included File Structure<\/strong>:<\/li>\n<li><code>pixelseal.pth<\/code>: Weights for the PixelSeal embedder and extractor (model states only) &#8211; <code>pixelseal.yaml<\/code>: Model card configuration defining the architecture, 256-bit payload, and 256-pixel processing size &#8211; <code>videoseal_attenuation.yaml<\/code>: JND (Just Noticeable Difference) attenuation channel map referenced by the configuration card &#8211; <code>audioseal_generator.pth<\/code>: 16-bit generator weights for AudioSeal (for embedding) &#8211; <code>audioseal_detector.pth<\/code>: 16-bit detector weights for AudioSeal<\/li>\n<li><strong>Video Processing Specifications<\/strong>:<\/li>\n<li>Target format: SDR H.264 outputs (video watermarking is skipped for HDR outputs) &#8211; Processing unit during detection: Extracts 1 frame out of every 8 decoded frames, resizes to 256&#215;256 pixels, and inputs them into the extractor. Bit logit values are averaged across all sampled frames &#8211; Detection criteria: Detected when <code>log10(p_value) &lt; -6<\/code>, which is the logarithm of the p-value based on a binomial test, is met<\/li>\n<li><strong>Audio Processing Specifications<\/strong>:<\/li>\n<li>Format and sampling rate: Targets stereo PCM audio, operating at native 24kHz or 48kHz rates (no resampling is performed) &#8211; Detection process: Scans the entire audio track, processing each stereo channel as a separate mono signal &#8211; Detection criteria: Detected when the watermarking probability exceeds 0.5 in more than half of the samples<\/li>\n<li><strong>License Terms<\/strong>: Released under the MIT license (<code>LICENSE-videoseal<\/code> and <code>LICENSE-audioseal<\/code> are included as upstream licenses)<\/li>\n<\/ul>\n<h2>Performance and Quality<\/h2>\n<p>The model card publishes benchmarks regarding the time required for detection processing and operating speed, as well as verification results on robustness from related papers.<\/p>\n<p>Detection processing can be executed on both CPU and GPU, but there is a significant difference in processing speed. In the publisher&#8217;s test environment, video detection times for 5 to 10-second clips at up to 4K resolution (excluding file decoding time) are reported as follows:<\/p>\n<ul>\n<li><strong>B200 GPU (single)<\/strong>: Approximately 15 to 30 milliseconds<\/li>\n<li><strong>CPU<\/strong>: 0.6 to 3 seconds<\/li>\n<\/ul>\n<p>Thus, utilizing a GPU environment allows for extremely fast watermark detection on video frames. Mechanisms are also in place for audio to scan full tracks and evaluate watermarking probabilities for each channel.<\/p>\n<p>Furthermore, according to the underlying technology paper (arXiv:2412.09492), the VideoSeal framework jointly trains the embedder and extractor while applying transformations such as video codecs in between, resulting in higher robustness compared to strong baselines even under severe distortion conditions combining compression and geometric transformations.<\/p>\n<p>On the other hand, the model card clearly notes the following limitations regarding quality and reliability. First, because this watermark is not a cryptographic signature, the failure to detect one does not allow one to conclude that the content is non-AI generated. Furthermore, since the model itself is based on a general-purpose model published by Meta, anyone can perform detection if they obtain the published weights and code, while keeping in mind that the watermark signal could potentially be weakened or removed by a sufficiently skilled attacker.<\/p>\n<h2>Strengths and Use Cases<\/h2>\n<p>Lightricks\/LTX-2-Marking is specialized for embedding and subsequently verifying invisible electronic watermarks to identify the source of content without compromising quality within video and audio generation and editing workflows using LTX-2.<\/p>\n<p>Specifically, it supports the following use cases and pipeline processes:<\/p>\n<ul>\n<li><strong>Watermarking in Video Generation and Editing CLI<\/strong>: Watermarks can be applied as an option during video generation or video editing. Watermark embedding is opt-in, and if not specified, outputs can be generated without altering the original pixel data or PCM audio data at all.<\/li>\n<li><strong>Individual Processing According to Pipeline<\/strong>:<\/li>\n<li><strong>HDR Generation<\/strong>: Video watermarking is skipped, and AudioSeal watermarking is applied only to the audio track. &#8211; <strong>Audio-to-video<\/strong>: Since the input source audio is passed through as-is, PixelSeal watermarking is performed only on the generated video frames. &#8211; <strong>Text-to-audio<\/strong>: Since there is no video, only the audio marker is applied.<\/li>\n<li><strong>Fast Artifact Verification<\/strong>: By utilizing the published detection scripts and model weights, users can directly determine in a local environment whether target MP4 files contain LTX-2-derived watermark signals.<\/li>\n<\/ul>\n<p>This pack is suitable for engineers who want to ensure traceability for media outputs from generative models or manage and identify generated content in their own operational environments.<\/p>\n<p><!-- lmw:files --><\/p>\n<h2>Distributed Files<\/h2>\n<p><em>Weight files published in <a href=\"https:\/\/huggingface.co\/Lightricks\/LTX-2-Marking\/tree\/main\">Lightricks\/LTX-2-Marking<\/a>, listed by this site from the Hugging Face API. Sizes are the actual file sizes.<\/em><\/p>\n<div class=\"lmw-table-scroll\" tabindex=\"0\" style=\"overflow-x:auto;-webkit-overflow-scrolling:touch;max-width:100%;\">\n<table style=\"width:max-content;min-width:100%;border-collapse:collapse;\">\n<thead>\n<tr>\n<th>File<\/th>\n<th>Size<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td><code>audioseal_detector.pth<\/code><\/td>\n<td>35MB<\/td>\n<\/tr>\n<tr>\n<td><code>audioseal_generator.pth<\/code><\/td>\n<td>59MB<\/td>\n<\/tr>\n<tr>\n<td><code>pixelseal.pth<\/code><\/td>\n<td>309MB<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p><!-- \/lmw:files --><\/p>\n<h2>How to Get It<\/h2>\n<p>This repository contains only model weights and configuration files; the execution code for embedding and detection is provided in the official LTX-2 repository (<code>packages\/ltx-pipelines\/src\/ltx_pipelines\/marking<\/code>). Since the LTX-2 library does not automatically download model weights, users must manually acquire them in their local environment. Because it is not a gated model, it can be downloaded directly without individual agreement to terms of use.<\/p>\n<h3>1. Installing Dependent Packages<\/h3>\n<p>Clone the LTX-2 repository and install it by specifying the additional dependencies for watermarking.<\/p>\n<pre><code class=\"language-bash\">uv sync --extra videoseal --extra audioseal\n# Or when using pip\npip install 'ltx-pipelines[videoseal,audioseal]'\n<\/code><\/pre>\n<h3>2. Downloading and Verifying the Weight Pack<\/h3>\n<p>Download the model files to your working environment directory using the Hugging Face CLI.<\/p>\n<pre><code class=\"language-bash\">hf download Lightricks\/LTX-2-Marking --local-dir path\/to\/marking\n<\/code><\/pre>\n<p>The pack&#8217;s loader, <code>create_markers<\/code>, assumes that all 5 types of model files exist within the same specified directory. Additionally, since <code>.pth<\/code> files are in PyTorch&#8217;s pickle format, it is recommended to verify that they match the SHA-256 checksum values provided by the publisher after downloading.<\/p>\n<pre><code class=\"language-bash\">sha256sum pixelseal.pth pixelseal.yaml videoseal_attenuation.yaml audioseal_generator.pth audioseal_detector.pth\n<\/code><\/pre>\n<h3>3. Usage in Generation Pipelines<\/h3>\n<p>When executing the video generation or editing CLI, passing the download directory path to the <code>--marker-root-path<\/code> argument automatically embeds watermarks into the output file. If this argument is omitted, watermarking is not performed.<\/p>\n<pre><code class=\"language-bash\">python -m ltx_pipelines.distilled \\\n... \\\n    --marker-root-path path\/to\/marking \\\n    --output-path out.mp4\n<\/code><\/pre>\n<h3>4. Detection Execution Code<\/h3>\n<p>To detect watermarks from generated media, call the API of the <code>ltx_pipelines.marking<\/code> module. Detection processing can be executed on either CPU or CUDA devices.<\/p>\n<pre><code class=\"language-python\">import torch\nfrom ltx_pipelines.marking import create_markers, detect_audio_mark, detect_video_mark\n\n# Initialize markers\nvideo_marker, audio_marker = create_markers(&quot;path\/to\/marking&quot;, torch.device(&quot;cuda&quot;))\n\n# Execute video and audio detection\nprint(detect_video_mark(video_marker, &quot;out.mp4&quot;))\nprint(detect_audio_mark(audio_marker, &quot;out.mp4&quot;))\n<\/code><\/pre>\n<p>Executing each detection function returns a <code>MarkDetectResult<\/code> object containing the <code>detected<\/code> flag, bit accuracy (<code>bit_accuracy<\/code>), <code>p_value<\/code> from the binomial test, and the decoded <code>payload<\/code>.<\/p>\n<p><!-- lmw:runnability --><\/p>\n<h2>Can You Run It Locally?<\/h2>\n<p>The publisher distributes this model as PyTorch (pth).<\/p>\n<p><strong>License \u2014 <code>mit<\/code> (Commercial use allowed):<\/strong> Permits commercial use, modification and redistribution, provided the copyright notice and license text are retained.<\/p>\n<p><em>Compiled by this site&#8217;s code from the published formats and the license field. License summaries are not legal advice \u2014 check the publisher&#8217;s original terms before relying on them.<\/em><\/p>\n<p><!-- \/lmw:runnability --><\/p>\n<p><!-- lmw:related --><\/p>\n<h2>Related Articles<\/h2>\n<ul>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/10\/01\/ltx-25-sdr-to-hdr-ic-lora\/\">LTX-2.5-22b-IC-LoRA-SDR-To-HDR Video Generation Model: File List<\/a><\/li>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/09\/30\/lightricks-ltx-2-5-22b-ic-lora-restore\/\">LTX-2.5-22b-IC-LoRA-Restore Video Generation Model: File List<\/a><\/li>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/09\/11\/lightricks-releases-lora-ic-lora-for-ltx-25\/\">Lightricks Releases Multiple LoRA and IC-LoRA Adapters for LTX-2.5<\/a><\/li>\n<\/ul>\n<p><!-- \/lmw:related --><\/p>\n<p><!-- lmw:next-steps --><\/p>\n<h2>What to Read Next<\/h2>\n<ul>\n<li><strong>Learn about the publisher<\/strong> \u2192 <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/publisher-lightricks-en\/\">Lightricks (LTX): models, licenses and articles<\/a><\/li>\n<\/ul>\n<p><!-- \/lmw:next-steps --><\/p>\n<h2>Sources<\/h2>\n<ul>\n<li><a href=\"https:\/\/huggingface.co\/Lightricks\/LTX-2-Marking\">Lightricks\/LTX-2-Marking (Hugging Face)<\/a><\/li>\n<li><a href=\"https:\/\/arxiv.org\/abs\/2412.09492\">Video Seal: Open and Efficient Video Watermarking (arXiv:2412.09492)<\/a><\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>Learn about Lightricks\/LTX-2-Marking, model weights and configuration files for invisible watermarking in the LTX-2 video and audio generation pipeline.<\/p>\n","protected":false},"author":1,"featured_media":14093,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[203],"tags":[4325,957,4327,4329,4331,4333,1547,4335],"class_list":["post-14094","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-image-video-and-audio","tag-audioseal-en","tag-lightricks-en","tag-lightricks-ltx-2-marking-en","tag-ltx-2-en","tag-ltx-2-marking-en","tag-pixelseal-en","tag-verified","tag-watermarking-en"],"lang":"en","translations":{"en":14094,"ja":14092},"_links":{"self":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/14094","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/comments?post=14094"}],"version-history":[{"count":1,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/14094\/revisions"}],"predecessor-version":[{"id":16985,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/14094\/revisions\/16985"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media\/14093"}],"wp:attachment":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media?parent=14094"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/categories?post=14094"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/tags?post=14094"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}