[Source]https://bbci.magich5page.com/magic/eco/runtime/release/645907a43592e704dcc222be?
The rapid development of artificial intelligence technology has brought more possibilities to Internet industry. Especially in the field of content creation, generative artificial intelligence technology has lowered the threshold for creation, enriched the Internet content ecology, and brought new changes and opportunities in information production and dissemination. But at the same time, the content generated by artificial intelligence is difficult to identify, and it also brings problems such as false information and infringement.
The "Provisions on the Administration of Deep Synthesis Internet Information Services" clearly require that deep synthesis service providers employ technical measures to attach labels to information content produced or edited using their services which do not impact their use; and where services have functions that generate or notably change information content, a conspicuous label shall be placed in a reasonable position or location on information content that is generated or edited, to alert the public of the synthesis and to avoid confusing or misleading the public.
Douyin has always been committed to encouraging expression and stimulating creativity. In order to provide a better community environment, we are publishing these watermark and metadata specifications in accordance with the "Douyin Platform Norms and Industry Initiatives on Content Generated by Artificial Intelligence", which we hope will help users and the industry to better use generative artificial intelligence technology within a compliant and reasonable scope.
Watermark Label Specifications: Aim to help AI generation tools and creators to use a unified style and position to alert users that the content is generated by AI, and minimize the users' negative visual experience caused by different styles of watermark.
Metadata labeling specification: Achieving industry-wide identification by standardizing the metadata format of content generated by artificial intelligence, and writing information into the metadata of related images and videos.
Scope of Application
In principle, the watermark and metadata specifications are applicable to the creation tools and content described below; for content generated by creation tools that are within the scope described below, it is necessary, in principle, to embed watermark and metadata information on the artificial intelligence content generation in the content production stage.
Scope of creation tools: Creation tools that provide generative artificial intelligence service functions, such as the generation of images or videos from text, etc.
Scope of content types: Images, video, or audio generated by artificial intelligence.
Normative reference documents
GB/T 38548.4-2020Digital processing of content resources - Part 4: Metadata
GB/T 7408-2005 Data elements and interchange formats, Information exchanges, Date and time representation
GY/T 259-2012 Next Generation Broadcasting Network (NGB) Video-on-Demand System Metadata Specifications
WH/T 51-2012 Image Metadata Specifications
WH/T 62-2014 Audio Asset Metadata Specifications
WH/T 63-2014 Video Asset Metadata Specifications
Terms and Definitions
AI generation: Content generated by artificial intelligence, which refers to videos, images, text, and other content that is automatically generated using artificial intelligence technology.
Watermark: Labeling by affixing an image that can be recognized by human senses to images, videos, or audios. Specific identification information such as product logos and user IDs can usually be used in images and videos.
Metadata: Data describing data, descriptive information about data and information resources.
Watermark Labels for AI-generated Content: Watermark information that is perceivable by humans and indicates that the multimedia content is artificial intelligence-generated,
Metadata Labels for AI-generated Content: Information embedded in the metadata of a multimedia file that is used to indicate that the multimedia content is generated by artificial intelligence.
Specifications for Watermark Labels for AI-generated Content:
| TEXT | AI Generated |
| Style |  |
| Watermark Size | 4*20@1x * Instructions for use: @1x represents a mobile canvas based on 375*812. If you export a larger canvas, please enlarge the watermark proportionally. |
| Whether it can be customized | Not-customizable |
Usage Norms 【Click to view source images】
Specifications for metadata labels of AI-generated content
Metadata labels for AI-generated content are information embedded in the metadata of a multimedia file, which is used to indicate that the multimedia content was generated by artificial intelligence. Related artificial intelligence creation tools can add metadata labels to the metadata of multimedia files when generating multimedia content; this metadata label can be recognized by other tools and platforms, meaning that different platforms and tools can mutually recognize metadata labels.
Metadata Definitions
| # | ATTRIBUTE NAME | ATTRIBUTE DEFINITION | Type | Length | Constraint | EXAMPLE |
| 1 | AIGC 标检 （AIGCLABEL） | To mark the content as AI-generated, using aigc as the mark. | string | 4 bytes | essential | |
| 2 | 版本 （ Version） | Version of the metadata specifications · The initial version is 1.0 | string | 3 bytes | optional | Version 1.0 |
| 3 | 生成工具（GeneratingTool） | The tools used to generate the AIGC content; may include the name of the product and company that generated the content. | string | 32 bytes | essential | Generating Tool: Douyin_Jianying |
| 4 | 时间戳 （Timestamp） | The local timestamp for the generation of the AIGC label with the time format of YYYY-MM-DDThh:mm:ss. The time stamp may be different from the creation times and publication times in the file metadata. | string | 19 bytes | essential | Timestamp: 2022-11-30T00:00:00 |
| 5 | 内容实体标识 （ContentID） | The ID of the AIGC content generated by the generation tool, including a combination of English letters and numbers. The ID has the function of uniquely identifying the content, and within a company, information related to the content can usually be obtained based on the ContentID. | string | 32 bytes | optional | ContentID: v020049a0000bdfsppk81uksdg2ok310 |
| 6 | 扩展信息 （ExtendInfo） | Self-definable extended information, such as custom signing of content and/or metadata to enhance security. | string | optional |
Expansion principle: The metadata label of Ai-generated content can support extension elements; the extension principle refers to the extension principle in metadata specifications WH/T 51-2012, WH/T 62-2014, and WH/T 63-2014.
- In the existing metadata specifications, if there is no appropriate element that can be reused, it is permissible to extend the element by itself.
- Self-extended elements cannot have any semantic duplication with existing elements.
- Newly added elements shall take precedence over elements in other metadata standards.
- If newly added elements reuse elements from other metadata standards, the source must be explained, and its semantics must be strictly followed when used.
In view of the wide variety of multimedia file formats and in order to ensure compatibility, these specifications provide two means of embedding metadata labels into file metadata:
- Add an extension field aigc to the file metadata, and write the identification metadata as the value of the aigc field.
- Write metadata labels into the existing comment field value of file metadata.
In the first method, the metadata label has a separate field in the file metadata and will not be confused with other fields, but there is a possibility that some file formats do not support custom fields. The metadata specifications WH/T 51-2012, WH/T 62-2014, and WH/T 63-2014 all contain comment fields, so the second method has universal applicability, but mixing with other content in the comment field may occur. In a specific application, an appropriate embedding method can be selected according to the file format. So long as compatibility is ensured, these specifications recommend using the first method.
Embedding method 1: Add an extension field aigc to the file metadata
| Encoding format | JSON format aigc: {“GeneratingTool”: value1, “Timestamp”: value2} |
| video example | ffmpeg -i input.mp4 -map_metadata -1 -c copy -movflags +use_metadata_tags -metadata aigc='{“GeneratingTool”:”Douyin_Jianying”,”Timestamp”:”2023-04-18T00:00:00″,”ContentID”:”v0300fg10000cf0kbc3c77ub10123450″}’ -f mp4 -y demo.mp4 |
Image (XMP metadata) example | img.modify_xmp({‘Xmp.dc.aigc’: ‘{“GeneratingTool”:”Douyin_Jianying”,”Timestamp”:”2023-04-18T00:00:00″,”ContentID”:”v0300fg10000cf0kbc3c77ub10123450″}’}) |
| audio example | ffmpeg -i ./6956245659821279268.mp3 -movflags +use_metadata_tags -metadata aigc='{“GeneratingTool”:”Douyin_Jianying”,”Timestamp”:”2023-04-18T12:00:00″,”ContentID”:”v0300fg10000cf0kbc3c77ub10123456″}’ -y ./test.mp3 |
Embedding method 2: Embedding the comment field in the file metadata
| Encoding format | JSON format comment: aigc:{“GeneratingTool”: value1, “Timestamp”: value2} |
| video example | ffmpeg -i input.mp4 -map_metadata -1 -c copy -movflags +use_metadata_tags -metadata comment=’aigc:{“GeneratingTool”:”Douyin_Jianying”,”Timestamp”:”2023-04-18T00:00:00″,”ContentID”:”v0300fg10000cf0kbc3c77ub10123450″}’ -f mp4 -y demo.mp4 |
| image example | exiftool -comment=’aigc:{“GeneratingTool”:”Douyin_Jianying”,”Timestamp”:”2023-04-18T00:00:00″,”ContentID”:”v0300fg10000cf0kbc3c77ub10123450″}’ test.png |
| audio example | ffmpeg -i ./6956245659821279268.mp3 -metadata comment=’aigc:{“GeneratingTool”:”Douyin_Jianying”,”Timestamp”:”2023-04-18T12:00:00″,”ContentID”:”v0300fg10000cf0kbc3c77ub10123456″}’ -y ./test.mp3 |
1. Why is Douyin publishing watermark and metadata specifications for labeling AI-generated content?
Article 16 of the "Provisions on the Administration of Deep Synthesis Internet Information Services" clearly requires that deep synthesis service providers employ technical measures to attach labels to information generated or edited by using their services that do not impact the use of the information. Article 17 further requires that in specific scenarios deep-synthesis service providers and users shall place conspicuous labels in reasonable positions and locations of generated or edited information content, to alert the public of the synthesis and avoid public confusion or misunderstanding. In order to ensure that the content generated by artificial intelligence is effectively labeled, and to minimize bad user viewing experiences caused by differing watermark styles, Douyin has released these identification watermark and metadata specifications.
2. What will happen after pictures and videos whose metadata information is written in accordance with the metadata specifications are published to Douyin?
When metadata is written into information content in accordance with the metadata labeling specifications, Douyin will notify users that the content is AI-generated in the required scenarios, including but not limited to adding a notification label that "this content is AI-generated" in information flows, video detail pages, etc.
3. If the produced content has already added an AI-generation watermark, do I still need to re-add one according to Douyin’s watermark labeling specifications?
We recommend that you add watermarks according to Douyin's label specifications to avoid inconsistent visibility caused by differing styles and positions of watermarks added by tools and creators, and to present the created content to users in the best style as much as possible.
4. What common file formats support the addition of aigc extension fields in metadata? What common file formats are not supported?
Common file formats that support the addition of aigc extension fields in metadata include image formats such as PNG/JPEG, video formats such as WMV/FLV/MP4/MOV, and audio formats such as MP3. Those that do not support it include HEIC image format, AVI/TS/ HLS video format, etc.
English
中文(简体)
Be First to Comment