Bowen is from The Temple of Convi
Qubits reports | Official account QbitAI
Now really everything can be JOJO!
Musk, who was already radiant, the next moment seemed to be directly "I am not a man!" ”
The mysterious and elegant smile of the world-famous Mona Lisa also seems to have become JO burning...
Another one of the same dimensions... Colonel, what are you doing, Colonel!
The above effects can be achieved by opening the web page Demo and clicking to upload any local image.
This suddenly attracted a large number of netizens to watch, not only Twitter heat 800+, online trial hugging face (Hugging Face) also lined up, a photo to wait at most four or five minutes.
Not only JOJO style, but also Disney style, League of Legends style... Enter either style of image to quickly apply that style to a new image:
Looking at this chuchu poor Kazilan's big eyes, and the innocent princess laughing, I just want to say... Old horse, quickly collect the magic power!
Jo make any portrait online
After watching the above demonstration, do you want to make a whole JO face yourself?
Let's try it out with the Developer's Hugging Face and Colab.
The first is the online version of Hug Face, which can be dropped in any local photo by clicking on the white space in the left box:
Click On Submit again, wait a dozen seconds... Standing in front of you is not someone else, but ko no muscle golden wheel big Sima Da!
This method can be played instantly by clicking on the link. However, although it is simple and fast, you will occasionally have to wait in line for a few minutes:
So strongly Amway Colab version, import any image into the test_input folder, such as we put a piece of ice here, and click Run:
Then continue running the following build module:
Jojo my ice goddess, this is also in your calculations GAN!
Moreover, colab also provides another way to play: import an image of any style and make a XX style generator yourself.
Well...... Isn't this the whole Tivat Continental version of Musk?
Upload a serious picture of an old horse:
Upload a proto-god-style portrait in the style_images folder:
(Uploading a two-dimensional style image may appear "can't find a face", you need to try a few more pictures)
Then fine-tune, wait a few minutes, and then click Run:
This resolute look, coupled with the posture of the old horse in the original picture overlooking the chest, feels that the next second can come to the sky!
Approximate style is obtained by GAN inversion
So, how does this method make it possible to learn the art style perfectly by referring to only one picture, and then apply it to other images?
Let's take a look at this model called JoJoGan.
It mainly obtains the approximate style through GAN inversion, and the main workflow is divided into four steps:
Invert the reference style image into approximate paired training data by GAN to obtain the corresponding stylized code;
Generate a real face image based on stylized code and match it with a reference style image to form pairs of data as a paired training set;
Based on these pairs of training data, fine-tune the StyleGAN;
Generate new samples using the fine-tuned StyleGAN.
The developers say that this model pays great attention to the style details under zero supervision, and has good versatility in different styles, and can be easily generalized to images of other styles.
From the second dimension to the technical house
Developer Min Jin Chong is also an old acquaintance of ours, having previously worked on the two-dimensional wife generator:
Min Jin Chong graduated from the University of Illinois at Urbana-Champaign (UIUC) and continued to pursue his PhD in the fields of machine learning, computer vision and image generation.
He previously interned at Byte for 3 months and now works with two co-students to create a fashion shopping app called Style Space, which allows users to try out and buy products in a virtual space.
His mentor, David Forsyth, is a CV veteran who has collaborated with Jean Ponce on computer vision: A Modern Approach:
Hugging Face Online Demo:
https://huggingface.co/spaces/akhaliq/JoJoGAN
Colab Online Demo:
https://colab.research.google.com/github/mchong6/JoJoGAN/blob/main/stylize.ipynb#scrollTo=LCLWiXoXwcJb
Thesis Link:
https://arxiv.org/abs/2112.11641
Reference Links:
[1]https://twitter.com/ak92501/status/1473522187491590148
[2]https://github.com/mchong6/JoJoGAN