Japanese actor Genu Hoshino has announced his marriage to Niigaki Noyi, who has been dubbed his “wife” by many male fans, in a statement released by his office yesterday.
Original: HyperAI hypernerve
Key words: generative adversarial network StyleGAN
“Wife married”, “star wild source took his wife pain”, “ye green knot”…… Hoshino source and Niigaki Knot clothes official declared marriage, many netizens issued the above exclamation.
There is also a wave of netizens in the open acceptance of the “brokenhearted” status quo, turned to care about The children of Niigaki Noyi and Hoshino Source, for fear that the children do not look like their mother.
Using an open-source model, BabyGAN, we predicted what Niigaki and Hoshino’s children would look like.
“River” is the name of their child in Escape Is Shameful but Useful.
According to BabyGAN’s predictions, if Niigaki and Hoshino’s baby had been a girl, the river of different ages might have looked like this:
BabyGAN generated giFs of her growing daughter
If the big river were a boy, it might look like this at different ages:
BabyGAN generated giFs of his growing son
What the hell is BabyGAN
BabyGAN is a Stylegan-based child appearance predictor that can generate or predict the appearance of future children after neural network processing based on encoder and generator, input father and mother images.
Prediction method: The neural network model based on GAN architecture is used to extract latent representation from the input parent image, and then the algorithm is used to mix it in a certain proportion to generate child image.
With latency direction, age, facial orientation, mood, and gender can be changed.
For the project address, visit Here
Encoders visit Here
This tutorial mainly demonstrates:
1. Load the trained BabyGAN model locally
2. Prepare images of both parents and obtain their latent representation
3. Model the child’s face
4. Adjust the gender, age and other parameters of the child to generate child images that meet the needs
Installation environment: Python: 3.6; TensorFlow: 1.15
Note: It is recommended that this tutorial be run on a GPU
The full tutorial can be found Here:
Detailed explanation of model training process
1. Preparation
2. Prepare parent images
3. Generate child images
4. Generate child images with certain features
The full tutorial can be found Here:
Stylegan-related highly acclaimed open source projects
The BabyGAN model is based on StyleGAN, and there are many other good open source projects based on StyleGAN and StyleGAN2.
StyleALAE
StyleALAE is an adversarial covert autoencoder based on the StyleGAN generator. It can not only generate 1024 x 1024 face images with the same image quality as StyleGAN, but also reconstruct and change face attributes based on real images at the same resolution.
The StyleALAE encoder uses the Instance Normalization (IN) layer to extract multi-scale stylistic information, which is combined into an implicit code W through a learnable multilinear map.
Related papers: Here
Project address: Here
StyleFlow
While it is easy to produce high-quality, diverse, lifelike images with StyleGAN, controlling the generation process with (semantic) attributes while maintaining high-quality output is not easy to achieve. In addition, due to the entanglement nature of GAN’s potential space, editing along one attribute can easily cause changes in other attributes.
In order to solve the problems of attribute-conditioned sampling and attribute-conditioned editing in the conditional exploration of entangled potential space, Researchers have come up with StyleFlow.
Non-sequential editing of real images with StyleFlow is better than concurrent method for extreme images such as old people and asymmetric images.
Related papers: Here
Project address: Here
Pixel2style2pixel (pSp)
PSp is a StyleGAN encoder for image-to-image conversion. Based on a novel encoding network, it can directly generate a series of style vectors, which are input into the pre-trained StyleGAN generator, forming an extended W + potential space.
In the pSp, the encoder can embed the real image directly into W + without additional optimization, and the image-to-image conversion task can be solved directly with the encoder, which is defined as the coding problem from the input field to the potential field.
PSp in StyleGAN inversion, multi-modal conditional image synthesis face positive, image repair and super resolution scene
The pSp can handle a variety of image conversion tasks, such as generating face images from segmented images, face-positive, super-resolution, etc., without changing its structure.
Related papers: Here
Project address: Here
GenForce
GenForce is an efficient PyTorch library for StyleGAN, StyleGAN2, PGGAN, etc. It has the following features:
1. Distributed training framework
2, training speed is fast
3, modular design, suitable for the prototyping of new models
4. Compared with the official TF version, StyleGAN training is highly reproduced
5, including many pre-trained GAN models with Colab Demo
Related papers: Here
Project address: Here
About OpenBayes
OpenBayes is a leading machine intelligence research institute in China, providing a number of basic services related to AI development, such as computation force container, automatic modeling, automatic parameter tuning, etc.
OpenBayes also has data sets, tutorials, models and other mainstream public resources available for developers to quickly learn and build ideal machine learning models.
Go to OpenBayes.com now and sign up
You can enjoy
600 min/week vGPU
And 300 minutes per week of free CPU computation
Go ahead and use BabyGAN to predict what your child will look like.
Complete tutorial portal: Here
Colab Portal: Here