digg

I'll be really impressed when the program starts making images of things no one has ever seen before ;)

You're going to be waiting for a little while. This algorithm used a few million high resolution images for semantic scene matching (for a grand total of about 400GB in source images) and a very complex algorithm for edge matching ("Local Context Matching" which requires two colorspace conversions per pixel, LAB->RGB and RGB->LAB, and an 80 pixel radius around the edge fragments). You could pre-generate a lot of the information and use a very huge hash table to help speed up the process, but it's still not a desktop application, and won't be for at least another 5 years (or another year or so if the graphics card companies change their business models and give us better access to the internals of their chips).

This kind of thing is better suited for Hollywood special effects departments; cut out some part of the frame, have the algorithm find a new image to fill in the blanks.

be nice if the info was posted on a better website with less spam

Baaaah, until I see it working with a click of my mouse I'll have serious doubts on this... ok, you can get similar colors, but "semantically valid" (i.e. recognize whats represented in a pic a match it with something appropriate)? I bet most of those pics were made with a lot of human "help"...

It's great, but as usual, the general public will probably never see this.

This is what Photoshop's vanishing point should have been.

Will this work with porn?

My guess is that shadows would fail to match a lot; I suppose the algorithm could be adapted to account for them.

Nothing will ever replace the human hand and eye when it comes to doctoring a photo, creating a photo illustration, or doing many of the other things you can do with Photoshop. What about people who paint in Photoshop, or create graphics for print media? You people are definitely underestimating humanity and Photoshop, and putting way too much faith in your ***** algorithms you babble about all the time.

It isn't as seemless at it those two examples make it seem. Look at the many examples given in the original research paper (http://graphics.cs.cmu.edu/projects/scene-completion/scene-completion.pdf warning 11mb pdf) and you will see that it more or less works well but does not produce perfect seamless images except in a few rare instances.

Then funny part is he's thinking of semen.

awesome thanks

most of those look really fake

The picture with the boats fits in really well so that you can't tell it was pasted in there. But to me the imagery inside of the block pasted in, the boats, doesn't seem to be too well thought out. The shadows from the boats appear to be cast in one direction while rest of the shadows in the picture form in the other direction. Not perfected yet but still amazing for such little human manipulation.

they're researching that.. It'll be like you were never here.

No. Because if you couldn't load a picture, how would it access a database?

"Our algorithm requires over an hour to process one input image on a single CPU (50 min for scene matching, 20 min for local context matching, and 4 min for compositing)."

Agreed. Gone are the days of those pictures you get from myspace friends with bad boobs. Just run em through this program after editing them out and voila, saggy 32B to perky 36C!

Yeah, agreed. Dugg for the Algorithm, not too sure about the application of it.

The PDF linked above hints at some useful applications. One I can think of is taking modern pictures of ancient cities, and using the algorithm to remove modern elements such as television antennas and power lines.

dugg down for lol usage.

My favorite tech blog is spunkrocketcumblaster.com

not really