Shulou(Shulou.com)11/24 Report--

While we are not paying attention, AI painters have been quietly iterating, and the recently launched Stable Diffusion Reimagine and Midjourney v5 are so powerful that they not only have to eliminate human painters, but also prompt engineers to lose their jobs.

This time, the human painter is really going to lose his job.

Can you believe that this picture like the stills in the movie was painted by AI?

What's even worse about the light, shadow and skin effects generated by Midjourney v5 is that AI painters use their own brains.

Recently, the new Stable Diffusion Reimagine, as long as you give it a picture, it will infinitely develop its own imagination. It will generate as many editions as you want.

With this AI, no matter how strange the request is and how many times it needs to be changed, the designer and painter will dump the plan that Party A's father wants to him in a few minutes.

Can you zoom in while the logo is enlarged? Change!

Want colorful black? Change!

He is so calm and at ease.

The only thing to worry about is that Party A's father must not know.

AI learns to imagine that humans are going to lose their jobs as prompt? Stable Diffusion is an AI drawing tool that we are all familiar with.

Yesterday, Stability AI revealed another shocking news: it has greatly improved Stable Diffusion.

This new product is called Stable Diffusion Reimagine.

Dare to call Reimagine, it sounds very shocking.

Yes, now all you have to do is upload one picture and SDR can create countless new images based on it.

And it is a real imagination, a real creation, not copying the faces or objects in the original picture, but re-creating a new picture based on the inspiration of the original image.

Experience address: https://clipdrop.co/ stable-diffusion-reimagine We all know that in previous AI drawing tools, the quality of prompt often determines the upper limit of how many pictures you can draw.

Many people say that whether or not prompt will become a watershed between mediocrity and genius who use AI. Similarly, prompt engineers see prompt as their secret weapon to open the AI award.

The creator of the Space Opera House, the winner of last year's art competition at the Colorado Fair, refused to share the tips he used on Midjourney. It is said that it took him more than 80 hours to complete the painting in 900 iterations.

Now, there is no need for complicated prompt, just throw a picture to AI, and the algorithm can create infinitely many changes we want.

Click the mouse and your bedroom will change immediately.

Technical principle this brand-new Stable Diffusion Reimagine is based on a new algorithm created by stability.ai.

The classic Stable Diffusion model is trained to be conditional on text input.

In Reimagine, the original text encoder is replaced by an image encoder. Instead of generating an image based on text input, an image is generated from the image. After the encoder passes the algorithm, some noise is added to produce changes.

This method produces images with similar appearance but different details and composition.

Unlike the image-to-image algorithm, the original image is fully encoded first, which means that the generator does not use any of the pixels from the original image.

In addition, Clipdrop can also improve the resolution of images. Users only need to upload a small image to get a picture with at least twice the level of detail.

It is reported that Stable Diffusion Reimagine's model will soon be open source on StabilityAI's GitHub.

Of course, Reimagine also has some limitations.

The biggest drawback is that it doesn't create images based on the original prompt.

In addition, its performance is not so stable, sometimes it is amazing, sometimes it is more stretched.

In addition, although Stability AI installs a filter in the model to block inappropriate requests, it is not a panacea.

In addition, the AI model can not avoid bias, Reimagine can not avoid, Stability AI is currently collecting user feedback, hoping to reduce the bias.

Netizens have gone crazy and of course they can't wait to play.

The upper left corner (or far left) is the original file, while the other pictures are "re-imagined" after being inspired by the original file.

From the point of view of the shared works, the most prominent effect is indeed the design:

Stable Diffusion Reimagine can redesign the entire scene without changing the overall atmosphere.

It can be seen that Stable Diffusion Reimagine has performed very well in the handling of hand-painted works.

The netizen said that it understands the "feeling" quite well, and he likes the effect of this freedom.

In addition, Stable Diffusion Reimagine is also good at dealing with the two-dimensional role.

In the real scene, some netizens praised after trying their own selfie: "I bet you will never guess which photo is real." "

However, when the editor tries some "famous scenes", the result is as follows.

First of all, enter "I want all" of the "nine-grade sesame official":

Then ask Stable Diffusion to "re-imagine": (the result more or less kills Matt in it)

Instead of "will Smith slapped Kerry Locke in the face at the Oscars"?

Ah, this.

Maybe we opened it the wrong way. After all, the main purpose of Stability AI is to help designers get similar options with just one photo.

So, let's do a test with the cases they give.

Note that the following is an official demonstration:

And the following is the "reproduction" effect of the same picture.

Obviously, the images generated by Reimagine look far less realistic, and the proportions are strange.

After several attempts, we still don't get an image that looks completely real.

In this regard, some netizens also said after trying that there is still a big gap between the quality of pictures generated by Stable Diffusion Reimagin and Midjourney.

Because Midjourney can be optimized through a large amount of feedback generated by the community, which is difficult for Stable Diffusion to match.

AI painters are finally able to draw hands, and another powerful AI painter, Midjourney, is certainly not idle.

On Wednesday, Midjourney announced that we have upgraded to version 5!

Compared with previous versions, V5 version has higher image quality, more diversified output, wider style, seamless texture, wider aspect ratio, better image prompts, and wider dynamic range.

What about the effect of Midjourney V5? what AI artists say is "too lifelike, too perfect, so creepy." "

"it's as if a myopic suddenly put on his glasses-the picture suddenly becomes 4k. "

Netizens commented: because every time is too perfect, too amazing, and finally stopped secreting dopamine

It can be seen that Midjourney has been improving in terms of image details since its debut in March 2022. Version 3 was launched in August last year, version 4 in November, and version 5 in March this year.

Enter a prompt such as "A muscular barbarian holding a weapon next to a CRT TV, movie, 8K, studio lighting", and the images generated by v3, v4, and v5 are as follows--

Midjourney v3 (left), v4 (center) and v5 (right), but the most amazing promotion is the handling of manpower.

Nice!

It is well known that models such as Midjourney, Stable Diffusion and DALL-E have been trained by millions of human artists.

To build LAION-5B datasets, robots directed by AI researchers crawled billions of websites, including large image libraries of DeviantArt, ArtStation, Pinterest, Getty Images and so on, and collected millions of photos.

However, even so, the AI painter has not learned to draw.

Even if the person and the scene have reached the degree that it is difficult to tell the true from the false, as long as you zoom in on the hand, it is basically a matter of a second.

Netizens have said that most of the "beauty selfies" that never show their hands are the work of AI.

Why is that?

Stability AI's explanation is that in AI data sets, human hands are not as significant as faces; hands tend to be small in the original image and rarely appear in large form.

AI and Amelia Winger-Bearskin, a professor of art at the University of Florida, explain that AI don't really understand what a "hand" is and how it anatomically relates to the human body.

Wieland found that Midjourney v5 can draw hands with five fingers most of the time, rather than seven to 10 fingers.

And after AI can draw hands, any photos that appear on the Internet can't tell the difference between true and false.

If there is anything wrong with Midjourney V5, it is probably too lifelike and perfect, which deprives us of the pleasure of finding the best result after many attempts.

(the editor who tried many times yesterday and finally drew a perfect picture of Lin Daiyu in the words of Wen Xin said that he very much agreed.)

This is the slot machine effect (near-miss)-so close, that's what keeps us going.

Reference:

Https://arstechnica.com/information-technology/2023/03/ai-imager-midjourney-v5-stuns-with-photorealistic-images-and-5-fingered-hands/

Https://stability.ai/blog/stable-diffusion-reimagine

This article comes from the official account of Wechat: Xin Zhiyuan (ID:AI_era)

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.