This page gives a detailed breakdown of classes that are intended to describe characteristics of an image rather than classify the image's subject matter. Examples include whether the image is natural (i.e., photographic) or animated, whether the image includes text etc.
These classes are rarely used for moderation directly, but may be useful in conjunction with other moderation classes (e.g., to exclude animated/illustrated images and art) or as part of other image understanding workflows such as OCR. You can choose which of these classes to use, if any, based on your needs and policies.
This model head can be used to identify images that contain legible text. This includes both text that has been digitally overlayed and scene text appearing on objects, clothing, etc. in the original image. For these classes, numbers, characters or script in any language is classified as text. This also applies for animated or illustrated images such as art, cartoons, and drawings.
yes_text: the image contains legible text or numbers, such as:
- Words or numbers in any language or script
- Logos or watermarks that include words or numbers
- Overlay text such as memes or captioned images
- Text or numbers visible on clothing, signs, billboards/posters, objects, etc.
- Images where most text is blurry or too small to make out, but at least one character or word is legible
no_text: the image does not contain text or any identifiable text is not legible. Includes:
- Images with no overlay or scene text
- Images that include signs, license plates etc. that are far enough away that individual characters can't be made out
- Images that clearly contain text, but the text is out of focus, blurred, or occluded
This model head can be used to identify images where text has been digitally added to an original source image. It does not capture scene text or other types of digital overlay such as drawings, emojis, or stickers.
The yes_overlay_text class is a subset of the more general yes_text class. It can be used to identify memes and other types of captioned images but ignores scene text in the original source image.
yes_overlay_text: the image includes text that has been digitally added to a photographic image or digital illustration, such as:
- Memes with top and/or bottom text (including illustrated memes)
- Watermarks and timestamps
- Images captioned with text such as social media stories, posts, etc.
- Screenshots from videos with text overlay elements, such as commercials or livestreams
no_overlay_text: the image does not include digital overlay text, or the image is an animation or physical drawing. The following would not classified as yes_overlay_text:
- Stylized or composite images such as movie posters and billboards
- Cartoons and anime
- Infographics, diagrams, and maps
- Non-text overlays, such as stickers, emojis, or doodles
- Scene text on clothing, objects, etc. that is part of the original source image
This model head can be used to identify images that include babies and toddlers. Generally, children older than three years old will not be identified as toddlers for these classes.
This class is designed to supplement main moderation classes to ensure that young children are not present in adult or NSFW content.
Art, illustrations, and animations are automatically classified as no_child_present. Photos of photos, such as images where pictures of children are displayed in a room, are also considered negative for child presence.
yes_child_present: the image is a photograph that includes a toddler or baby. The child does not necessarily need to be fully in the frame (e.g., head, backside, legs, etc. clearly belonging to a child would also be flagged)
no_child_present: one or more of the following is true:
- The image does not include a baby or toddler; OR
- The image includes an older child or teenager
- The image is a composite showing a toddler or has a photograph of a toddler in the image
- The image is an illustration, drawing or animation of a baby or toddler
- The image includes a baby doll
This model head can be used to identify images that are drawings or illustrations (including photographs of the same).
yes_drawing: the image shows or contains:
- Drawings, sketches, and paintings
- Digital art
- Doodles and illustrations, including posters and album artwork if illustrated
- Comics, cartoons, animations, manga
- Textbook illustrations and diagrams
- Art, drawings and paintings displayed in the setting/background
no_drawing: the image does not include drawings, art, illustrations, etc., including:
- Sculptures, models, and similar 3D mediums
- Video game graphics
- Live action TV shows and movies
This model head can be used to identify whether an image is illustrated/animated, photographic, or a combination of photographic and illustrated elements.
Digital overlay text and annotations such as arrows or circles used to draw attention to a particular part of the image are not considered to be drawings or animation. These types of images will not be classified as hybrid.
animated: the image is purely illustrated, includes:
- Photographs of drawings, paintings, illustrations etc. whether they are the only visible element of the image
- Animations and cartoons
- Digital art and computer-generated designs
- Video game graphics
natural: photographic images that are not combined with illustrated elements:
- Photographs, both color and black and white
- Photographs of drawings and illustrations where other elements or objects are visible
- Photographs of paintings where the frame or other elements are visible
- Photographs of statues, models and other 3D art
- Composite images or collages where all elements are from photographs
- Photos with filters or effects that don't include a drawing or animated elements
hybrid: the image includes both photographic and illustrated/animated elements. For example:
- Color-manipulated images
- Photos modified to include drawings or animated elements such as cartoon characters, avatars etc.
- Photos with overlay stickers, graphics, or emojis
- Collages that include both drawings/paintings and photographs
- Photos with filters that include animated elements (e.g., cat ears)
Updated 11 days ago