There was a paper recently where a research team trained a machine learning algorithm (a GAN they called AttnGAN) to generate pictures based on written descriptions. It’s like Visual Chatbot in reverse. When it was just trained to generate pictures of birds, it did pretty well, actually.
(Although the description didn’t specify a beak and so it just… left it out.)
But when they trained the same algorithm on a huge and highly varied dataset, it had a lot more trouble generating a picture to go with that caption. Below, I give the same caption to a version of their algorithm that has been trained to generate everything from sheep to shopping centers. Cris Valenzuela wrapped their trained model in an entertaining demo that attempts to generate a picture for any caption.
This bird is less, um, recognizable. When the GAN has to draw *anything* I ask for, there’s just too much to keep track of – the problem’s too broad, and the algorithm spreads itself too thin. It doesn’t just have trouble with birds. A GAN that’s been trained just on celebrity faces will tend to produce photorealistic portraits. But this one, however…
In fact, it does a horrifying job with humans because it can never quite seem to get the number of orifices correct.
It’s fun to ask it to draw animals though. It knows the texture of giraffes, but not quite exactly their shape. And it knows that boats are on the water, but not necessarily that they are boats.
It also (like many other image recognition algorithms) gets a bit confused about the difference between sheep and the landscapes they’re found on. Other algorithms recognize sheep in pictures of empty green fields. And this one, when asked to draw sheep…
That’s different, though, from asking it to draw *a* sheep. In that case, it knows exactly what to do. It draws the sheep, and then just to be safe it fills the entire planet with wool too.
It really likes drawing stop signs and clocks. Give it the slightest opportunity to draw one, and it will chuck those things all over the place.
Other than its horrifying humans, this algorithm can actually be pretty delightful.
I had way too much fun generating these and ended up with way more than would fit in this one blog post. I’ve compiled a few more of my favorites. Enter your email and I’ll send you them (and if you want, you can get bonus material each time I post).
As someone not involved in the development of pillowfort but am a web developer, I think you should lower your expectations, but not for the reason you think.
Pillowfort is a baby. A newborn. A smol bab. If you were here during the early days of Tumblr, think of that.
Pillowfort simply cannot be the immediate solution to your woes. It needs to be nurtured and cared for to become a mature and happy adult.
If you want Pillowfort to work, they’ll need feedback, advice, bug reports, etc. This is a chance to make Pillowfort the Ao3 of Fanfiction.net. It’s not gonna happen overnight, you need to give it time and love and it’ll get there.
If you don’t want to pay money to get into the beta, that’s ok. It will be open to the public soon enough and you won’t have to pay a dime. Their financial model moving forward sounds good (a subscription fee for super extra features), but even an Ao3 model would work swell for them probably.
We’re living in an interesting time on the internet. Governments across the world are cracking down on content and yet community run websites are starting to thrive more and more.
Tumblr once upon a time was what Pillowfort is today, but this time, let’s make sure Pillowfort can stay independent from mega corporations.
yes this
most of the criticism i’ve seen of pf so far ultimately come back to this
is it an alright platform with a good community? yes. does it have a lot of potential? yes. does it have a lot of problems still being worked out? yes. are the staff open and responsive? absolutely. do i recommend it? yes, if you’re willing to live in a house while it’s being built.
but it’s not a ready-made replacement for tumblr. set your expectations accordingly.
So out of morbid curiosity I tried Tumblr’s export feature to download my blog. Pressed the button, waited. After a few hours I got an email it was ready.
For my relatively modest blog of ~5K posts/reblogs, it produced a zip file of about 12 GBs. It didn’t say how big the file was when downloading; I just had to wait until it was done. Once downloaded, Windows 10′s native zip management couldn’t handle it, insisting it was a broken archive. An ancient dusty install of 7Zip popped it right open, though.
Inside were two folders and one .xml file:
“Media” clocks in at 12 GBs and consists of 14K gifs, jpgs, pngs, mp3s, mp4s, and movs. “Posts” has a folder of individual stripped-down HTML docs of every post, plus an xml doc that also seems to be every post in a single 21MB file. (This doc does appear to include my 800 draft posts.) And messages is all your messages/chats (which I admit is nice to have a backup of, though the xml is of course unreadable without some kind of reader).
But the best, the BEST (sarcasm level 8) part of this is that all those HTML/XML post files? They link back to the files ON TUMBLR. They don’t have internal links to the files in that Media folder. So when you open one of your downloaded posts in your browser, the images you’re seeing are from Tumblr’s servers – for as long as those images are posted and you’re online.
So this reblogged post, ‘capped from my downloaded archive:
Meanwhile files in the Media folder are just numbered by the post ID number, not those strings. It would be possible, I’m guessing, to run a conversion, to switch all the media links to using the ID numbers to link to what’s in the media folder – but that’s beyond my limited regex skills. Without that, you’re left with a pile of media with no organization whatsoever, and a folder full of posts with broken media.
(Things like this make me wonder if they’re not trying to repackage the site for advertisers at all but are just trying to kill it quick. Then again, trying to ascribe any kind of firm rational motive to whoever’s in charge of this feels like accusing a clogged toilet of having an agenda…)
Thank you for doing this! (Also … OMG TUMBLR. >__>) To some extent that’s actually somewhat better than I was afraid of, because at least you have all the images and the text of the posts, both in easy to read formats; I was worried it’d be one giant horrible thing that you’d need some kind of decoder program to read, and you’d have to pull out images individually if you wanted them. (I mean, back in Strikethrough days I remember finding a 3rd-party site that would download your journal as a PDF but didn’t include most of the images or comments. This sounds like an improvement over that, at least!)
… but still.
Also, I started a backup running yesterday and nearly 24 hours later it’s still just showing me “backup processing” which is … not promising. (This file is going to be huuuuuuge.)
Oh yikes. Well, that answers my question anyway. Sheesh.
Some good news, I’ve been talking to two developers now and got them working together, we just had a meeting with the guys behind an existing large (millions of users) site similar to Tumblr, with a vibrant and open-minded community, and more importantly, it has open-minded owners who believe in free speech. They think we can get something done here to rescue the whole community.
I’m not allowed to reveal the site name yet. I can tell you it’s mainstream, open to everyone, open-minded and welcoming. (It’s not WordPress or any site owned by Facebook or Twitter. It’s not Pillowfort, that’s in closed beta. It’s not Ello, that’s mainly for artists. It’s not kinkspace or fetlife, those are too specialist. It’s not jux, that seems to be closed. It’s not Soup, that seems still in development and too small.)
One of the reasons for delaying the announcement for next few days is they don’t want a “land grab” where people take the names of current popular Tumblr users over there (cyber squatting). So they are looking at ways for existing Tumblr users to keep the same names on the new site.
More info over the days to come.
The plan is, broadly:
1. By December 9th, announcement of the new site and how to secure your username there
2. By December 10th, an online tool for bloggers to copy their existing content to the new site automatically, with the same tags and captions.
3. Bloggers will need to copy their content across between December 10th and December 17th if they want to use the automatic tool.
4. My understanding is that after December 17th there will be no public access to any “flagged” posts on Tumblr, but the original poster will still be able to see the flagged post (for a short time at least). Therefore, the original poster may still be able to manually download a post to their own PC or phone, after December 17th, and manually upload it to the other site. But if you have lots of posts that will take a long time, it will be better to use the automatic tool before December 17th.
Please understand that these dates are approximate and may change for technical or other reasons.
There may be a few rough edges or not so perfect looking site design on the transfer tool. Everyone is doing their best. The main goal here is to help as many people as possible preserve access to their content, in the short space of time Tumblr has allowed us, and preserve as much as possible of the Tumblr community spirit somewhere new.
The new site will cater for photo, GIF, text and html posts. It will not offer video and audio posts, due to cost reasons – maybe in future, but for now you will need to preserve video and audio content yourself in some other place.
If your Tumblr blog has a mixture of original content and reblogs, or all reblogs, all of that can be copied over to the new site. Reblogs will become “your” original content if nobody else posted them yet, otherwise they will be shown as reblogs. The devs are looking at ways to preserve attribution of reblogs back to the original Tumblr poster, if that person also moves to the new site.
Important: your Likes cannot be copied from Tumblr to the new site. You will have to go find the same posts again on the new site, and like them afresh.
(Similarly, existing reblog comments, asks, messages and other user interaction on Tumblr cannot be copied to the new site – that’s just too much to do, in the short time available.)
If you want to preserve any of your existing Liked posts on Tumblr, you will need to either: (1) download the post to your own PC, or: (2A) reblog it now to your own Tumblr blog, and then (2B) use the automatic tool, before December 17th, to move your whole Tumblr blog across to the new site.
If you have Liked a lot of posts here on Tumblr, the gridllr.com webapp should be able to help you do steps 1 and 2A quickly, I mean download or reblog.
(Someone complained to me today about the appearance of Gridllr on a phone. It’s best to use Gridllr on a PC, Mac or Tablet with a large screen.)
If you have liked a post here on Tumblr and the original poster decides to delete it, or even to delete their entire blog, some time before December 17th, then that post will be permanently lost. So if you want to be sure to preserve any of your Liked posts, you should best download or reblog as soon as possible. If it’s reblogged to your own blog it is safe from deletion, at least for next few days.
Obviously, you will lose access, after December 17th, to all past posts you have liked, if Tumblr has flagged them as NSFW. Again, the steps (1), or (2A) and (2B) covered above will be the only way to hold on to these posts.
lovingdread 