Jonathan Fly  

@jonathanfly

Using the wrong tool. The least efficient way. For no good reason. 👾 YouTube (twitter compression 😔)

Vrijeme pridruživanja: travanj 2009.

Tweetovi

Blokirali ste korisnika/cu @jonathanfly

Jeste li sigurni da želite vidjeti te tweetove? Time nećete deblokirati korisnika/cu @jonathanfly

  1. Prikvačeni tweet
    17. ruj 2019.

    Trying depth-mapping on abstract art. I don't know what you would call 'working properly' here, it's just fascinating to see a neural network try to interpret a Jackson Pollock painting as a 3D landscape. (Convergence, 1952)

    Prikaži ovu nit
    Poništi
  2. That's enough Keanu for one night on Twitter, but I swear he makes interpolation artifacts look cool, almost natural.

    Prikaži ovu nit
    Poništi
  3. Some transitional frames look like reflections: shots from a very different movie, full of pathos.

    Prikaži ovu nit
    Poništi
  4. Doesn't work as well for people only in the scene for a single frame. Star Trek episodes excepted.

    Prikaži ovu nit
    Poništi
  5. Make real video look like a GAN interpolation by motion smoothing up from 1 fps. The depth based geometry really works in a couple of places -- looks like a model rotating in a scene consisting entirely of 2 or 3 frames.

    Prikaži ovu nit
    Poništi
  6. My dream is that Twitter is saving all original video uploads and one day a switch is flipped and all old Tweets burst into HD. DAIN Windows port: Works on consumer GPUs, lots of memory needed for HD. YouTube of original sample:

    Prikaži ovu nit
    Poništi
  7. Some people asked about DAIN. It's Depth-Aware Frame Interpolation. I like to try low frame-rate sources like 16fps 8mm family films. Footage is supposedly anonymous but that sure looks like Stan Lee... project: Original Footage:

    Prikaži ovu nit
    Poništi
  8. 2. velj

    It may be easier to get intermediate poses if you don't try and keep the character fixed, though most of these still look more like growing rather than animating.

    Prikaži ovu nit
    Poništi
  9. 2. velj

    I can't keep the clothing fixed during this animation. According to the model, if you dual wield pistols, you wear a trench coat. Or at least a trench dress. Those are the rules/tangled-latent-vectors.

    Prikaži ovu nit
    Poništi
  10. 2. velj

    Artbreeder can almost animate characters by crossbreeding. I'm not sure if the model has intermediate poses -- it looks more like the limbs grow into position than move. Also testing DAIN motion smoothing.

    Prikaži ovu nit
    Poništi
  11. 2. velj

    Blue II is an exception: a huge cliff at 8x zoom. Strange, I don't see any depth cues? I tried swapping colors and I think it's as simple as blue = distant-blue-sky because "Red II" is completely flat. Not sure where the cliff border is from though. Blue II, Joan Miró, 1961

    Prikaži ovu nit
    Poništi
  12. 2. velj

    I never posted Ken Burn 3D abstract art samples because they come out mostly flat in the depth model. I just double-checked though and there are some geometry differences. I froze the frames in this video that feel like strange vector re-interpretation of the original work.

    Prikaži ovu nit
    Poništi
  13. 1. velj

    Ken Burns 3D has blob artifacts even in the first frame with non-centered zooms. I wonder if scene rotation could be delayed until a later zoom level? A few examples, first is centered. Patrick Street, Dublin City, 1898 Photo and texting-time-traveler spotted via

    Poništi
  14. 1. velj

    Friday Night Artbreeder Pirates. Seeing a lot more character variety across the site in the last 24 hours. Breeding two images is a really effective way of exploring latent space.

    Prikaži ovu nit
    Poništi
  15. 1. velj

    I'd love to see what this does with non-furniture images. Their doesn't output 3D but '' has various voxel-related conversions (npz, ply, rle) none of which I know a thing about. May be an easy tweak for anyone who works with voxels.

    Prikaži ovu nit
    Poništi
  16. 1. velj

    Semantic Scene Completion is inferring labels (what something is) and 3D geometry at the same time. PALnet: Depth Based Semantic Scene Completion with Position Importance Aware Loss abs: pdf: code/weights:

    Prikaži ovu nit
    Poništi
  17. 31. sij

    ~41 seconds slowed down and (badly) interpolated. Quite the jubilee. 🎉

    Prikaži ovu nit
    Poništi
  18. 31. sij

    This one isn't particularly good with the color or 3D mapping, but the geometry and artifacts "behind the image" are some of my favorites. Diamond Jubilee decorations, St James’s Street, London 1897, DeOldify, Ken Burns 3D.

    Prikaži ovu nit
    Poništi
  19. 31. sij

    GauGAN has a lot of frame-to-frame variance if you try to generate video, but that same variance makes for a lovely effect if you lean into it. A slow zoom on a fixed segmentation map becomes a breezy summer day.

    Prikaži ovu nit
    Poništi
  20. 31. sij

    NVIDIA's latest GauGAN vs last - feels like the 'camera' focuses differently? Hard to compare as style choice changes so much though. It isn't meant to render people but I appreciate that they let it try. I love the GauGAN 'no-people' so much I'm using one as my profile image.

    Prikaži ovu nit
    Poništi
  21. proslijedio/la je Tweet
    22. sij

    Here is the Colab notebook for Generation by Cooperation Over a Noisy Channel! [Warning: this code may be quick and dirty; it may have some future experiments hinted at as well ;) ]

    Prikaži ovu nit
    Poništi

Čini se da učitavanje traje već neko vrijeme.

Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.

    Možda bi vam se svidjelo i ovo:

    ·