Jul 21, 2022 19:00:00

'This Food Does Not Exist' that can generate high quality photos of food that does not exist in the world

Developed and released by Uber engineers in 2019, '

This Person Does Not Exist ' has attracted worldwide attention as a tool that can generate faces of non-existent people using AI. After that, ' This Anime Does Not Exist ' that generates animation images that do not exist in the world, ' This Foot Does Not Exist ' that was created for foot fetish, and ' This cat does not exist ' that can generate pictures of fictitious cats. Various similar tools have appeared. A new Swiss machine learning engineer, Florian Laurent , has announced the development of ' This Food Does Not Exist ,' which can generate pictures of food that doesn't exist in the world.

GitHub --nyx-ai / stylegan2-flax-tpu: Training StyleGAN2 on TPUs in JAX
https://github.com/nyx-ai/stylegan2-flax-tpu

This Food Does Not Exist | stylegan2-flax-tpu
https://nyx-ai.github.io/stylegan2-flax-tpu/

'This Person Does Not Exist', which generates a human face photo, used ' StyleGAN ' based on the hostile generation network (GAN) , which is a kind of AI algorithm, but 'This Food Does Not Exist' Is based on ' StyleGAN2 ' which is an improved version of StyleGAN.

In the machine learning industry in recent years, the diffusion model used for ' DALL-E ' developed by OpenAI has become popular, and the diffusion model is catching up with the StyleGAN model, which is superior in terms of photographic realism. thing. Laurent also points out that while the diffusion model has excellent controllability because it can generate images from text information, it requires a larger model and has a long generation time. It is explained that 'This Food Does Not Exist' was developed using StyleGAN2 this time because it costs a lot of money to use the diffusion model.

Laurent used one of Google's machine learning processors, the TPU v4-8 , to train each of the four StyleGAN2 models in less than 10 hours. This is the picture generated by 'This Food Does Not Exist' trained with the image of the cookie. The size of each image is 256 pixels x 256 pixels.

On the other hand, the image of 'a pile of cookies on a plate' generated by 'DALL-E' is now ...

This is the image generated by '

Craiyon ' derived from 'DALL-E'. The image generated by 'This Food Does Not Exist' is comparable to 'DALL-E', and it can be seen that it is more accurate than 'Craiyon'.

Besides cookies, Laurent also has cheesecake ...

Cocktails

I am training 'This Food Does Not Exist' to generate images such as sushi.

'This Food Does Not Exist' isn't published as a web service, but code and pre-trained models optimized for TPU training are published on GitHub.