I couldn’t get a picture of Trump and Biden at Five Guys, but I did get pictures of the Founding Fathers:
This is why I prefer Playground AI; they don’t have any restricted words. You could put in, “big fat juicy penis vagina asshole pussy cunt motherfucker jizz,” and it will run. You may not get a result, but it’ll give it a shot. Another plus is sometimes, if you’re running in SD 1.5, you may actually get an image depicting a big fat juicy penis vagina asshole pussy cunt motherfucker jizz, because there’s definitely porn trained in there.
That’s because they quietly delete the censored words in the background. I remember someone confused because their images of cucumbers were coming out as cubes, and it turned out to be that Playground was deleting “cum” and leaving behind “cuber”. There is a workaround for that, though: embed a “bad word” in your “bad word”. If you are aiming for a penis, write “pepenisnis” and it will delete the central “penis” and pass “penis” along to the AI.
Thanks.
We’re fortunate Ed Zotti slipped through the net.
I read something from six months ago saying Bing will save 20 sets of four images each, after which they are lost.
Is this correct?
If so, I presume images posted in the thread, and links to images posted on the thread, will become broken/lost?
You’re in the OR, about to be put under for abdominal surgery. A moment before the general anesthesia causes you to lose consciousness, you see your surgeon wheel out a tray of surgical instruments…
What 3D dataset would you train it on?
Also, a question: I want to experiment a bit with one or more of the models that can do photorealistic images. What’s the best one to sign up for, that doesn’t require payment, nor unjustified saleable personal information?
A variety of sources. There’s no shortage of 3D assets used for games, etc. with descriptions. As well as 3D printer assets from Thingiverse, etc.
I expect the training will use 2D as well, though. Feed it a million 2D human heads, and somewhere deep in the weights it will “know” how heads exist in 3D space. It has to figure this out if it’s to construct novel heads that look realistic. The trick will be to somehow connect this knowledge with something that can be converted to an STL file.
There is plenty of opportunity for 2D->3D conversion as well. So in principle, you could just connect two nets together to get a text->2D->3D pipeline. But ideally, if you train everything together, the net can eventually work around the lossiness you would normally get with a 2D intermediate step. It would instead have some more sophisticated intermediate representation that contains the details needed to eventually produce an STL file.
The images shared “offically” through the link on the Bing page are likely treated differently than those just scrolling off your timeline. But you should always download every file you have any interest in keeping.
For SDXL:
For Dall-E 3:
In the time that I’ve actually been able to use Dall-E 3 I have been making lots of snail-human hybrid images, which happened to be one of the first prompts to come to mind to try. The previous Bing Dall-E already made pretty interesting images with that concept, but DE3 makes some amazingly coherent and genuinely creepy images. I haven’t posted any because I like to outpaint images I like beyond the 1:1 aspect ratio using Clipdrop and Playground and I have been building up a stock of edited images.
What I want to ring up right now is nudity. Out of the 120ish sets of images I’ve made with DE3 so far, five of the images have included Unrequested Boobage. So we know there was at least female topless nudity in the training set, despite their wanting to keep things SFW.
Since these images aren’t s for w, home, or anywhere else, I’m hiding the link.
All seems safe for now. Interesting that there’s an image of a robot thinking about a robot painting an artwork.
“DALL·E 3, what are you thinking?”
If you have a half-decent PC, the best bet would be to run Stable Diffusion locally for consumer privacy plus image quantity limited only by your time and PC specs. There’s some easy-to-install clients out there and a ton of free models available. The underlying models many of these use (Stable Diffusion 1.5 or SDXL) is the same across the board.
Not my intent to fight the question so I tried to wait until someone else offered some online services first but it could be worth investigating.
It is hilarious how unprepared Microsoft was to handle the traffic from this. I’m still getting only sporadic assess and instead of mentioning a longer than normal delay they have a new popup that tells you they are too busy, come back later.
I’ve got three Bing assignments that are still waiting to be completed, the earliest from a couple of days ago, while there have been more recent assignments I’ve tasked that only took a couple of minutes to complete. I received one finished assignment around a day after I ordered it.
And you thought I wouldn’t get a result with my keywords. Don’t worry, it’s not that gross, but there is a pussy in there:
So I tried this prompt in Dall-E 3 and the results were… um… different. (This is a single output image.)
