The famous Song Dynasty painting “Tiger Wearing VR” has become popular on the Internet – yqqlm

Who would have thought that the tigers of the Song Dynasty would one day be played with fire abroad.

The thing is, Google didn’t come out with an AI creation artifact Imagen not long ago.As long as you give a sentence, it can generate semantic pictures.

Then foreign netizens with big brain holes gave Imagen a question without a routine:

Wear VR to the Oriental Tiger of the Song Dynasty.

Imagen is not afraid, “snap” gave a masterpiece – “Tiger Wear VR“.

Not to mention, this “Tiger Wearing VR” is really a bit of that flavor. Not only the style of painting, but the VR headset can be consistent with the tiger and the entire painting. Even the handle and the playful feeling of the two tigers are drawn in one step.

Then there are two tigers wearing VR, holding hands and “chacha” together:

Even Imagen has ingeniously designed a “connected” version of VR (maybe watching the film face-to-face):

But after all, in the matter of AI painting, there are many artifacts besides Google Imagen. As a result, a “Tiger Wear VR” painting battle kicked off.

(Guess whose painting is more like “Tigers in the heart, sniffing roses”)

DALL-E also comes to fight

The first to fight, must be DALL・E from OpenAI.

Netizen Jacob, out of curiosity, used it to make a few pictures for comparison.

The first is “Tiger Wearing VR” full of “fixed makeup photos” style (very sassy):

It is not difficult to see that DALL·E’s paintings and Imagen are still very different in style.

Imagen’s paintings tend to be more minimalist, while DALL·E is a little moreelements of oil painting.

However, in terms of artistic conception, DALL·E can also produce “two tigers at play” and even anthropomorphic paintings:

Compared with the two, netizens gave their evaluations:

Most netizens pay more for Google’s Imagen.

And apart from these two, like AI painting artifact MidJourney also participated in this “war”. But it’s a little bit weird…

DALL·E VS Imagen

So, as AI creation artifacts, why are the recent popular Imagen and DALL·E painting styles so different?

Open AI’s DALL·E and Google’s Imagen can both generate surreal-like images directly from text descriptions, allowing machines to have designer-like creativity.

but,The “creation” principle of the two is very different.

DALL·E 2 employs CLIP to map textual features to image features, and then directs a GAN or diffusion model to generate images.

The so-called CLIP is a neural network trained on various images and texts, sorting multiple generated images, and selecting better generated results for display.

Google’s Imagen uses a pure language model that is only responsible for encoding text features, leaving the work of text-to-image conversion to the image generation model.

The language model part uses Google’s own T5-XXL encoder, which freezes the trained text.

The image generation part is a series of diffusion models, which first generate low-resolution images and then supersample them step by step.

Google’s T5-XXL has 4.6 billion parameters, and expanding the scale of the text encoder can effectively improve the text-to-image correspondence and image fidelity.

In addition, Imagen uses another diffusion technique called noise conditioning augmentation, which helps the model learn the amount of noise that has been added to improve the restoration of the image.

By comparison,Imagen seems to be more “realistic” than DALL·E:

At present, all kinds of novel images have emerged on the Imagen official website.

Someone put an astronaut helmet on a raccoon.

The teddy bear starts swimming butterfly here.

There’s also eagle-shaped chocolate ice cream (well, that’s fitting).

As of now, Imagen and DALL·E are still in the debugging stage,Not yet open to the public.

One More Thing

In this AI painting battle of “Tiger Wearing VR”, there are also some failed works.

For example, some netizens gave an example of using DALL·E mini to generate.

It is not difficult to see that in this version of “Tiger Wear VR”, there is no VR, and the tiger’s face is basically blurred.

According to the description of netizens, in the process of generation, he just changed “Northern Song Dynasty” to “Southern Song Dynasty”:

The most difficult “figurativeness” of paintings has declined this time.

So which AI artifact do you think is stronger in “Tiger Wearing VR”?

Reference link:

https://twitter.com/hardmaru/status/1532757753797586944?s=21&t=MhwVN5VXH22zFK7DWQJnCg