Just a guy jumping from a hot mess into more prosperous waters.

  • 1 Post
  • 73 Comments
Joined 1 year ago
cake
Cake day: June 22nd, 2023

help-circle










  • That’s fine, but not the primary issue.

    At some point these companies will need to get licenses for any copyrighted work that was part of the training data, or start over with public domain works only. The art may be data, but that data has legal owners whose rights grant control over it’s use.

    Another way to think about is proprietary code. You can see it and learn from it at your leisure. But to use it commercially requires a license, one that clearly defines what can and cannot be done with it, as well as fair compensation.


  • The short version is that it’s a licensing issue. All art is free to view, but the moment you try to integrate it into a commercial product/service you’ll owe someone money unless the artist is given fair compensation in some other form.

    For example, artists agree to provide a usage license to popular art sites to host and display their works. That license does not transfer to the guy/company scraping portfolios to fuel their AI. Unfortunately, as we can see from the article, AI may be able to generate but it still lacks imagination and inspiration; traits fundamental to creating truly derivative works. When money exchanges hands that denies the artist compensation because the work was never licensed and they are excluded from their portion of the sale.

    Another example: I am a photographer uploading my images to a stock image site. As part of ToS I agree to provide a license to host, display, and relicense to buyers on my behalf. The stock site now offers an AI that create new images based on its portfolio. The catch is that all attributed works result in a monetary payment to the artists. When buyers license AI generated works based on my images I get a percentage of the sale. The stock site is legally compliant because it has a license to use my work, and I receive fair compensation when the images are used. The cycle is complete.

    It gets trickier in practice, but licensing and compensation is the crux of the matter.


  • A lot of nuance will be missed without some gradation between “I <3 China” and “Down with Pooh!” For example, if we added “Slightly favorable”, “Neutral”, and “Slightly unfavorable” we would begin to see just how favorable younger generations are. Rather than presume there is a deep divide on trade policy, if two bars are almost equal, we may see they are largely neutral. Similarly we could see just how favorable their views of TikTok really are by looking at the spread between neutral to “I <3 China!”




  • The short version is that there are two images and sidecar/xmp file sandwiched into one file. First is the standard dynamic range image, what you’d expect to see from a jpeg. Second is the gain map, an image whose contents include details outside of SDR. The sidecar/xmp file has instructions on how to blend the two images together to create a consistent HDR image across displays.

    So its HDR-ish enough for the average person. I like this solution, especially after seeing the hellscape that is DSLR raw format support.