I just got access to DALL·E 2 and here are my first few results
currentlyobsessed.comI am amazed by level of comprehension that DALL-E shows. At first glance the results also look amazing. Zoomed in though it turns into the stuff of nightmares. Quite literally actually. I am an enthusiastic practicionor of lucid dreaming and this stuff really feels similar to what I see when I closly observe details in lucid dreams. Ostensibly everything looks real but this reality falls apart when actually observed.
I too noticed how "dream-like" it is. Like it seems to have similar capabilities and limitations to how my brain renders things in dreams... they look fine at a glance but if you focus on details things get weird.
Now that I think about it I have always had great trouble with reading text in my lucid dreams. Dalle is also bad at making comprehensible passages.
Like what? I went back and looked at all of them closely again. Maybe the only weird thing is the mummies faces. Other than that it really is stellar. I mean just look at the detail on the “shaped like a heart” ice cream scoop, or the paper mache Godzilla.
1 What is on the pizza? is it a pizza at all?
2 whats up with bush and rocks in the middle of the street?
3 sculpture of a bird (beak) and a dolphin?
4 volcano erupting snow instead of smoke/lava?
5 what are those glasses?
6 actually pretty good as its an abstract object
7 teeth made out of meat
8 abstract = good?
9 bottom left oyster meat
10 grill grate geometry got really wonky
Every single picture looks very aesthetically pleasing, but gets reality wrong. Its all a very good content aware fill.
Look closely at the cone, fingers, or the people behind the ice cream.
Or the food in the last one.
The detail breaks down.
The ash above Mt Rainier looks more like mountains.
The thumb in the cone pic looks odd.
The food looks fine, so does the ash over the mountain.
All the fingers in the pictures look rather off and also the left eye (edit: left in the picture) of the dog looked rather scary.
To say the images fall apart is quite pedantic IMO.
These images would easily pass as hyperrealist paintings.
The content of the images also is quite tame if you can basically tell it anything.
FYI, openai has eased the rule on realistic face generation. Now you can generate and publish Photorealistic faces. They will internally filter those to make sure they don't match famous faces.
I had the weirdest reaction to that last sentence.
As an engineer, the first thing that went through my head was "lots of photos of their faces available, easy filter. My face specifically: hard filter"
But as a person it was "ok wow, f_ck you too then?"
So the famous faces everyone can easily find online anyway: those need protection. Gotcha.
I assume it's more so you can't do something like "Hilary Clinton stabs a child in a dark basement caught on webcam" and then pass it off as fake news somewhere.
How would DALL-E 2 be more effective than plain old photoshop in that case?
Using Photoshop has a much higher skill barrier than typing a fake news headline.
I mean yeah but it just costs money to hire talent, only difference with DALLE is that its subsidized to be free
I would love to see what comes out with certain aspects of the prompts negated.
- "lemon gelato that’s been shaped to look like a heart, on a handmade waffle cone being held up to the camera in a cobblestone courtyard somewhere in italy" ... what about "somewhere not in italy"?
- "Diorama made of clay of a group of computer programmers looking disapprovingly at their CMO who has just given them diet pepsi instead of mountain dew" ... "looking approvingly"?
- "friends gathering around a tabletop “shichirin” grill where an assortment of meats and seafoods are being grilled over glowing binchotan charcoal; everyone is happy." ... "everyone is unhappy"?
Ok, here are some results:
1: "not in italy": https://ipfs.io/ipfs/QmWPbZnZL7mHazzMmGxx6wQjcbC2DdKtgpiYYUY...
2: "looking approvingly": https://ipfs.io/ipfs/QmNgD9niZy1n4KWSXS3HEFeZm2qxQ2SzESfA6z7...
3: "everyone is unhappy": https://ipfs.io/ipfs/QmcTktpFQGeGDAp7e3MwEPwwFMVMbDaEgW6PuSA...
I think getting good Dall-E results wil end up being an "art" in its own. Dall-E is like a broad brush and honestly I've never been good at getting great results. I think figuring out how to push Dall-E in a way that aligns with what you want with the right descriptors really goes a long way.
I think to get there, we need a good dictionary or wall of examples that tell you what you can even do. I didn't even know you could have it create clay dioramas.
It's interesting that the gelato's colour seems to have become the wall colour as well. What happens if you ask for strawberry gelato, or lime?
Looks like a big weakness of DALL-E 2 is mixing up the properties of every object and of the background/setting.
https://twitter.com/david_madras/status/1512573390896480267
https://www.lesswrong.com/posts/uKp6tBFStnsvrot5t/what-dall-...
Mew
the way it generates specific art styles and textures is amazing. it's interesting to try and spot out subtle details that it misses/ignores entirely (who is drinking diet coke out of a mug? why are the mummies interpreted as skeletons in sock-raincoats?)
How long were you on the waiting list before you got access? I only signed up a week or two ago, so I assume I'm in for a long wait.
> How long were you on the waiting list before you got access? I only signed up a week or two ago, so I assume I'm in for a long wait.
It's not based on duration, but rather multi-faceted criteria which it seemingly determines based on email association, social media, Linkedin, Geographical location etc... it's not quite clear how it's weighed but those are all things it asks you for when you resister for the Beta [0].
The thing that was found out recently that their might be a paid public release coming soon because they integrated Stripe into Dall-E 2 [1].
0: https://www.reddit.com/r/dalle2/comments/vpind8/i_received_a...
1: https://www.reddit.com/r/dalle2/comments/vo16pj/a_billingcre...
Well, that puts me to the back of the queue then. I don't do social media or LinkedIn>It's not based on duration, but rather multi-faceted criteria ... social media, Linkedin... etc
I'm not the OP but I signed up ASAP in like April or whenever it was first announced and I haven't got in yet.
I signed up in January and just got access a couple of weeks ago
How much fine-tuning and repeated-running did you do to get these? Some of them are just ridiculously awesome.
Dall-e2 is the most amazing technology I have ever seen. Literally magic.
Something on this site tries to access my camera. I can hear it resetting when the page loads.
probably an analytics plug-in trying to fingerprint you, I remember i've had a lot of trouble with websites interrupting audio on my phone when there's not anything playing
I'm curious what kind of camera you have that makes a sound when accessed
On the Pinephone the rear camera homes the mobile into one end of the track when it's reset and makes a clicking sound.