Apple releases Depth Pro, an AI model that rewrites the rules of 3D vision
109 points by bentocorp 13 hours ago | 20 comments
  • sva_ 12 hours ago |
  • fh973 12 hours ago |
    This article has a link to the live demo.

    https://huggingface.co/spaces/akhaliq/depth-pro

    For some pictures it outputs something reasonable, for others it's completely broken (black with colored noise in one area).

  • ipsum2 12 hours ago |
    Title is such clickbait, it does not rewrite the rules of 3d vision, it is a marginal improvement on existing models, and does not work for video, only images. However, Apple open sourced the model weights, which is amazing for research.
    • Tepix 11 hours ago |
      Right, the title is so awful that i didn't even feel like reading the article.
      • esperent 10 hours ago |
        I made it about a third of the way down. It doesn't get any better. Gave up when I hit the auto playing unrelated video that you can't scroll past. Do people really keep reading an article while a video about something else is playing on the top third of their screen? Totally nuts.
      • bsenftner 8 hours ago |
        It's AI generated PR, perhaps the worst use of AI.
    • fxtentacle 10 hours ago |
      Intel open sourced Midas, too, which seems to have pretty similar result quality.
    • bamboozled 9 hours ago |
      “The greatest model, just got better, introducing…” can I have a job now ?
    • amelius 9 hours ago |
      What is the topology of the model like?
  • LeoPanthera 12 hours ago |
    This presumably is the same model that the Vision Pro Photos app uses to convert 2D photos to 3D.
  • dyauspitr 11 hours ago |
    Can I use this to generate accurate depth maps from 2-D images that I can then CNC or 3-D print?
    • j5155 9 hours ago |
      No. You’ll want a photogrammetry or LIDAR app for that.
    • stavros 8 hours ago |
      Maybe, if you're OK with printing just the front face of an object with its depth being roughly estimated.
      • pavlov 8 hours ago |
        You can of course do some great art with just a depth map manifested in a substance like marble:

        https://en.wikipedia.org/wiki/Parthenon_Frieze

        (Exaggerating a bit. Ancient Greek reliefs do have sculpted detail on the underside, e.g. the horse legs come slightly detached from the surface. So they are not 100% depth maps.)

    • KaiserPro 7 hours ago |
      Kinda, but you dont get the back.
  • zimpenfish 11 hours ago |
    Just tried it on a "difficult" image (relatively low contrast photo of a small thin plant in front of a tree trunk with a distant fence in one corner) and it did a pretty good job, I think - https://imgur.com/a/Sqr6hR8 including the depth maps.
  • kylehotchkiss 11 hours ago |
    Was this trained on iPhone photos since there is a decent amount of depth references within iPhone cameras? It’s interesting to see how clearly it understands depth of field. With that, how does it perform on F16 and above?
  • skykooler 9 hours ago |
    interesting. They claim 0.3 seconds on a consumer GPU; I thought that might scale to 30 seconds or so on CPU but gave up waiting after twelve minutes.