My understanding is that the 'truncation' parameter makes the output of BigGAN more closely resemble the training data set.
Moving the truncation towards 0 will give you more realistic but less varied images, while moving it towards 1 will give you more varied but increasingly potentially nonsensical images.
> Combine sugar and sugar and sprinkle with the sugar.
Seems about right describing current food lol
On a more serious note I wonder if our grand-grandchildren will look at these programs from blog posts and see them as some kind of "PRINT HELLO; GOTO 10" for AI, ancient artifacts showing the birth of a new paradigm of automation
An earlier entry from the same blog on generated recipes. My wife and I were laughing so hard at these we got some weird looks from the waiter. The use of asterisks is priceless in that one.
Funny how they mess with your brain: the illusion is that you are looking at something familiar. The texture is "easy" to look at, but the brain fails to recognize a known object.
Interesting article, but the subject of the article (the amazing images) are only shown in tiny thumbnails, with no links to the full-size images. Seems to sort of defeat the purpose.
Even the largest images produced by the network are only 512px x 512px and most are smaller — adding more pixels to these kinds of networks is fairly expensive and not at all worth it if you're training thousands of them over and over again while doing research.
The train picture reminds me of some of my botched attempts at creating perspective by combing two differing horizon line sub-images into one. I suppose that sort of rendering abstraction is the same in terms of accidentally producing weirdness - even though different processes were used.
We should have a reverse image search built in image-net for someone doubting and trying to find similar images in imagenet . Like google's reverse image search.
Very doubtful, the results just seem to be combination of existing background and existing artifacts (nose, ears, furs, etc) .
The different poses you would get naturally is much more complicated and diverse than these networks seem to be generating. All in all these networks aren't generating anything new.
The images you saw in the article were hand picked to be weird. Most of them are photorealistic and are not copies of any training photo. If you think it's so easy, paint a photorealistic image in Photoshop by hand (no direct copying, but you can look at other images all you want) and let's see how it compares.
If you are competing with a human, if not me I am sure a professional can easily better the best of best (and not just 512x512 pixels but 100 time more pixels) by this network. It has to catch up with humans by a long shot.
> Sure, human face comes in all shapes and sizes, so its tricky to judge the effectiveness since every modification is a possible real person.
What I wanted to show you is that you can walk through the 'latent space' of faces and generate intermediary images between any two images, thus they are not simple copies.
I think that that will depend on the license and the courts. A NN can be seen as a storage system - these images are retrieved from the compressed and transformed data stored in the NN; a court might rule that that is essentially the same as any other storage system. Or not.
Examples from the notebook by the article author: https://twitter.com/JanelleCShane/status/1062067001504321536