Search results
  1. 18 Jun 2019

    We have released our under the BSD license (It was under CC non-commercial license.) Please feel free to use it.

  2. Jan 21

    Recently I put together a small dataset of Bentley's classic snowflake photos, and trained a model. Here I generated some shapes with Pts.js and ran them through the model. Initial results -- (I probably need to train it some more but it's so expensive! 😭)

  3. 15 Mar 2019

    Glad to see that our research works enable people to "generate realistic dance videos of NBA players for in-game entertainment." ,

    Show this thread
  4. I've created an experimental GAN architecture I call or "Recursive-Residual GAN" and I am pretty astonished that: - it works at all - how well it works across a pretty wide range of scales. - it is just 15% the size of a comparable model

    Portrait generated by RecuResGAN - replication of training example
    Portrait generated by RecuResGAN - replication of training example
    Portrait generated by RecuResGAN - replication of training example
    Portrait generated by RecuResGAN - replication of training example
    Show this thread
  5. 1 Dec 2017
  6. Research at the lab | turning video games into interactive mode to question the value of designed pixels in future graphical productions | This is controlled with a body

    Show this thread
  7. by "Synthesizing and manipulating 2048x1024 images with conditional GANs"

    Show this thread
  8. 27 Aug 2018

    Hopping on the frame prediction train. Here's a feedback loop between 2 models, one trained to predict the next frame in a video, the other trained to predict the original frame from a version processed by the other model between 1 and 10 times.

    Show this thread
  9. 7 Nov 2018
  10. 19 Dec 2019

    Some generated snow crystals from GAN, based on Wilson Bentley's classic photos. Training is still in progress, slowly slowly. More to come soon!

    Show this thread
  11. Both models are shallow ResNets derived from . I first tried UNets, but there the models learned to cheat very quickly and just abused the first skip connection to pass the information almost uncompressed.

    Show this thread
  12. By the end of 2017 my efforts to improve resolution were obliterated by two major breakthroughs: in short succession first showed their highly realistic celebrities made with and followed up with shortly after.

    Show this thread
  13. Jan 24

    1. Draw shapes in , by hand 2. Run with , fingers crossed 🤞

  14. The principle is pretty simple: in a classic residual architecture you chain several residual blocks behind each other (in the default is 9 blocks), what I do in is to use a single block, but loop 9 times over it, feeding its output back into its input.

    Show this thread
  15. I personally think if this image is converted into segment ( how pix2pixhd works ) then i m sure var won’t make these types of errors. We need image to outline based concept here.

  16. 12 Apr 2018
  17. 5 Feb 2019

    Some more initial results from a series that extends work exhibited at . Using streetview data, we trained a model to transform depthmap images to photographic images.

    Show this thread
  18. 22 Dec 2017

    My first test of having a neural network trying to recreate a scene using pix2pixHD. The test data is from captured frames from a game I made in Unity, where each object is corresponding to a specific label. Enjoy!

  19. Replying to

    These are some snapshots from the training of my custom version of using on the Costică Acsinte archive.

  20. Replying to

    Thanks! There are 5 different GANs involved which employ my own architecture that owes a lot to and .

Loading seems to be taking a while.

Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.