5 ms·

These enhance effects typically use guesswork. That would be useful to gain hints, but not to know for sure what the original looked like.

by jasfi 2y ago

thih9 2y ago
Unless the rest of the image offers details that the human observer would ignore but the AI would pick up. E.g. low res photo from 1828 of a US president - enhance to John Quincy Adams. Surprisingly there's a daguerreotype of John Quincy Adams from 1843[1], the AI could use that; perhaps "enhance" the time backwards a bit, to make him 15 years younger. [1]: https://en.m.wikipedia.org/wiki/File:John_Quincy_Adams_-_copy_of_1843_Philip_Haas_Daguerreotype.jpg https://en.m.wikipedia.org/wiki/File:John_Quincy_Adams_-_cop...
- itishappy 2y ago
  Right, and that would be guesswork unless the image had been explicitly tagged.
  - pavel_lishin 2y ago
    There are search engines that search for similar images; a small photo of John Quincy Adams may be 99.99% identifiable as such.
    - itishappy 2y ago
      An educated guess then, maybe even a good one! What happens if it's only 80% sure? Do you get a high-res John Quincy Adams or an admission that it can't identify the image content? If I feed it a low-res picture of Michael Faraday or Henry Clay, will I receive the finely sculpted jowls of John Quincy Adams in return?
  - thih9 2y ago
    Guesswork, deduction, tomayto, tomahto ;)
- pavel_lishin 2y ago
  Isn't that what Samsung was doing with photos of the moon? https://www.theverge.com/2023/3/13/23637401/samsung-fake-moon-photos-ai-galaxy-s21-s23-ultra https://www.theverge.com/2023/3/13/23637401/samsung-fake-moo... But I also had a similar thought. Use a combination of an image search engine like TinEye, and something like ClearviewAI, to actually either find a larger image of the submitted photo, or of more photos of the person in question, and reconstruct what they look like based on all of that additional data.
  - thih9 2y ago
    Yes. But I'd like to take it one step further. E.g. if I send a photo of an olympic podium where winners are pixelated and ask the AI to enhance, I'd like it to use other elements of the pixelated photo to recognize the olympic event and year, search some other database to find the winners and enhance the photo so that the people on the podium were those who actually stood on that podium.
derefr 2y ago
The one exception being video-based image super-resolution techniques, where the frames of the video are composited together to gain more information than is in any individual frame (which is also how MRIs work!) Tangent to that: does anyone know if there are any free-to-use tools (or open-source models) for doing video-based image super-resolution? (I imagine that such a tool could take video files directly; but that the actual underlying model would likely require that you first convert your video into an image containing a grid of sampled frames, making it essentially an image2image model that learns to sample from reference images. A bit like the image2image models trained to output depth maps given stereo-camera photo pairs.)

These enhance effects typically use guesswork. That would be useful to gain hints, but not to know for sure what the original looked like.

by jasfi 2y ago

These enhance effects typically use guesswork. That would be useful to gain hints, but not to know for sure what the original looked like.

thih9 2y ago

Unless the rest of the image offers details that the human observer would ignore but the AI would pick up. E.g. low res photo from 1828 of a US president - enhance to John Quincy Adams. Surprisingly there's a daguerreotype of John Quincy Adams from 1843[1], the AI could use that; perhaps "enhance" the time backwards a bit, to make him 15 years younger. [1]: https://en.m.wikipedia.org/wiki/File:John_Quincy_Adams_-_copy_of_1843_Philip_Haas_Daguerreotype.jpg https://en.m.wikipedia.org/wiki/File:John_Quincy_Adams_-_cop...

itishappy 2y ago
Right, and that would be guesswork unless the image had been explicitly tagged.
- pavel_lishin 2y ago
  There are search engines that search for similar images; a small photo of John Quincy Adams may be 99.99% identifiable as such.
  - itishappy 2y ago
    An educated guess then, maybe even a good one! What happens if it's only 80% sure? Do you get a high-res John Quincy Adams or an admission that it can't identify the image content? If I feed it a low-res picture of Michael Faraday or Henry Clay, will I receive the finely sculpted jowls of John Quincy Adams in return?
- thih9 2y ago
  Guesswork, deduction, tomayto, tomahto ;)
pavel_lishin 2y ago
Isn't that what Samsung was doing with photos of the moon? https://www.theverge.com/2023/3/13/23637401/samsung-fake-moon-photos-ai-galaxy-s21-s23-ultra https://www.theverge.com/2023/3/13/23637401/samsung-fake-moo... But I also had a similar thought. Use a combination of an image search engine like TinEye, and something like ClearviewAI, to actually either find a larger image of the submitted photo, or of more photos of the person in question, and reconstruct what they look like based on all of that additional data.
- thih9 2y ago
  Yes. But I'd like to take it one step further. E.g. if I send a photo of an olympic podium where winners are pixelated and ask the AI to enhance, I'd like it to use other elements of the pixelated photo to recognize the olympic event and year, search some other database to find the winners and enhance the photo so that the people on the podium were those who actually stood on that podium.

derefr 2y ago

The one exception being video-based image super-resolution techniques, where the frames of the video are composited together to gain more information than is in any individual frame (which is also how MRIs work!) Tangent to that: does anyone know if there are any free-to-use tools (or open-source models) for doing video-based image super-resolution? (I imagine that such a tool could take video files directly; but that the actual underlying model would likely require that you first convert your video into an image containing a grid of sampled frames, making it essentially an image2image model that learns to sample from reference images. A bit like the image2image models trained to output depth maps given stereo-camera photo pairs.)