Datawhale dry information
Latest: DALL·E 3 , Editor: Qubits
It’s amazing, OpenAI actually merged AI painting and ChatGPT !
No, the newly released DALL·E 3 directly brought two major shocks to AIGC——
The prompt word threshold is greatly reduced
Detailed descriptions that understand the nuances of semantic meaning are incredibly accurate.
That's right, the new version of DALL·E 3 not only eliminates the need for prompt word engineering, but also improves language understanding capabilities to a new level!
Just imagine it. For words , ChatGPT helps you expand; for drawings , DALL·E 3 gives you precise details.
AI can be like filming a movie, ensuring that every detail from the background to the characters is reproduced verbatim:
Under the full moon , the streets are bustling with pedestrians enjoying the bustling nightlife.
At a corner stall, a young woman with fiery red hair and a signature velvet cloak was haggling with a grumpy old vendor.
The grumpy hawker, tall and sophisticated, wearing a neat suit and a striking mustache , was chatting animatedly on his steampunk phone .
In addition to basic details, DALL·E 3 can even interpret vague adjectives such as prosperous, bargaining, and grumpy to life, which is no longer on the same level as CLIP.
At the same time, compared to the previous generation of old models, DALL·E 3’s own painting skills have also taken a big step forward:
Such an operation immediately stunned netizens.
Some netizens have decided to cancel their Midjourney subscription. "If Midjourney can't understand the text accurately, it's not even a competition."
Some netizens ridiculed that this is simply putting pressure on the upcoming Google Gemini.
For more details, let’s take a look at the effects displayed by DALL·E 3 one by one.
Use it directly in ChatGPT
Compared with the previous two generations, the biggest advantage of DALL·E 3 is that it is natively built on ChatGPT .
Not only does it mean a huge leap in language understanding, but even prompt words can be written by ChatGPT itself .
More details are hidden in the promotional video where Ultraman can’t help but boast about his cuteness.
This is the story of a parent who turned a five-year-old's fantasy into reality .
First, parents asked ChatGPT, "My 5-year-old baby has been talking about a 'super sunflower hedgehog'. What should it look like?"
You can see that ChatGPT has written four prompt words in different styles at the same time and given the corresponding images.
After parents choose one of the illustrations with a fairy tale style, the image of the protagonist of the story, the little hedgehog, seems to be fixed , and they can continue to ask ChatGPT to draw more.
By the way, give the little hedgehog a name Larry, so that you don't have to say "Super Sunflower Hedgehog" every time in subsequent conversations.
The protagonist has been decided, and then more elements will be added to make the entire fairy tale richer, such as drawing a house for Larry.
This not only demonstrates the ability of DALL·E 3 to create a consistent image , but also shows that LARRY's name is correctly written on the mailbox, which solves the problem of the previous version of DALL·E not being able to write .
Since we already use ChatGPT, why not improve the storyline?
The plot you just compiled will have matching illustrations available immediately.
Keep the character image and move to a completely different sticker style without any problem, you can print it out directly.
Pay attention, here comes the most amazing part, just let ChatGPT summarize all the content in the previous conversation and write it into a complete bedtime story.
Although the demonstration ends here, it is completely conceivable that with the ChatGPT plug-in function, an e-book can be directly generated.
It reminds me of the previous suggestion by netizens that the best way to ensure the safety of AI is to have employees of OpenAI, Anthropic and other companies have children.
This way they have an incentive to make sure the world is safe when AGI arrives. (Manual dog head)
Although DALL·E 3 cannot be played immediately, you can still take a look at the large number of samples released at once.
When you click on each card, you can also see prompt words, which are described directly in human words without adding complicated spells.
The combination of complex scenes and non-existent concepts has a stunning effect.
When making interior design concept drawings, the relationship between light and shadow cannot be faulted at first glance.
Coupled with the ability to write correctly, it is also very productive to post a poster directly (there are still some problems with the small fonts where the text is not specified).
Noam Brown, the father of poker AI who just joined OpenAI, also posted pictures of his trial robot playing cards.
CEO Altman’s favorite picture is “Avocado Seeing a Doctor.”
Some netizens tried the effect of using the same prompt words on DALL·E 2. They could only say that the words were wrong, the hole in the middle of the avocado had no words, and the treatment was even worse...
Do you still remember when DALL·E 1 was first released in January 2021? The place where the dream begins is a set of "avocado sofas".
No wonder netizens lamented: Look how far it has come!
"If there is any infringement, please delete the picture."
Of course, in addition to the above features, OpenAI also previewed some magical new features.
For example, DALL·E 3 will soon be equipped with an image discriminator .
This classifier can help identify whether an image was generated by DALL·E 3, not only to avoid accidental injuries (manual dog head), but also to quickly claim it as your own when DALL·E creates a good work.
As for generating images, OpenAI said it has done a lot of work to prevent it from generating violent, pornographic or other harmful images, or images with the names of public figures (stars, celebrities, etc.).
Regarding privacy, the New York Times previously broke the news that OpenAI is using certain technologies to blur faces in images uploaded to ChatGPT.
This is also to prevent ChatGPT from becoming a complete "face recognition tool", especially for celebrities whose photos have been circulated on the Internet.
Now this technology may also be used in DALL·E 3 to prevent the generation of infringing images .
At the same time, OpenAI is also working with the security red team to improve its image risk assessment capabilities.
In addition, in terms of training data, OpenAI also learned to protect itself this time with Midjourney’s “lessons learned from the past”.
Rather than going to court directly with the artist or waiting to be sued, OpenAI released a training data “disclaimer” on its official website :
You can disable our web crawler GPTBot from accessing your website by filling out the form. Alternatively, you can send an image that you want to keep private and we will remove it from the training data.
However, some netizens are not satisfied with the DALL·E 3 demonstration effect, thinking that it is not as good as the pictures produced by senior MidJourney players. OpenAI's funds are many times abundant.
Some netizens turned on Leeuwenhoek mode and began to pick out the details of the prompt words missed in the demonstration picture one by one.
For example, this cup is missing miniature lightning bolts.
There are only cannon wreckage scattered on the seabed here, but no treasure.
Whether these problems can be improved by adjusting the prompt words will not be known until you actually play the game.
So when will DALL·E 3 be launched? Highlights:
ChatGPT Pro membership ($20/month) and Enterprise edition available in October .
A standalone version will be available later this fall (currently $15 for 115 plays).
Reference links:
[1]https://openai.com/dall-e-3
[2]https://www.nytimes.com/2023/07/18/technology/openai-chatgpt-facial-recognition.html
[ 3]https://twitter.com/sama/status/1704547625482203560
Good stuff to learn, like three times in a row ↓