Great idea to add model provenance.
I'm not sure if Dalle-2 embeds any metadata, but openAI also has this CLIP model https://openai.com/blog/clip/ that lets you get embeddings. So we will be using this for our search functionality. It works amazingly well. Stay tuned for the search functionality. It is super powerful for prompt engineering.
Thanks for pointing out the copyright status. Most of the images in Discover are crawled on the web, and you can see all the sources under each image. However for the images that people post on our platform, we should definitely add the clarification for copyright status.
Hah! Same word, two totally different meaning.
Metadata embedding: store extra information in non-visible fields in the file. For example some phones save the gps location where the picture was made in the image.
CLIP embedding: a vector which describes the semantic content of an image. It is the output of an AI model which maps the image into a semantic space (aka embeds the image in said space.) This vectorial description then typically can be used for image content classification or textual label generation.
I feel the DALL-E 2 watermark signature could be a seed or something.
Adding model provenance would be a good idea so you can support multiple. I was wondering if DALLE 2 embeds any metadata in the images but haven't had a chance to investigate.
I also think you should consider adding the option to clarify copyright status ... in particular, allow users to dedicate the image to the public domain via CC0 (https://creativecommons.org/share-your-work/public-domain/cc...), or maybe put it under a CC BY license. May be debatable what the legal status of the images is, but a CC0 dedication would at least get rid of any ambiguity.