Style-Mixing in StyleGAN/StyleGAN 2

user_14247168 0
Viewed 1917

I have been training StyleGAN and StyleGAN2 and want to try STYLE-MIX using real people images. As per official repo, they use column and row seed range to generate stylemix of random images as given below -

Example of style mixing

python run_generator.py style-mixing-example --network=gdrive:networks/stylegan2-ffhq-config-f.pkl \
  --row-seeds=85,100,75,458,1500 --col-seeds=55,821,1789,293 --truncation-psi=1.0

I have below questions where I am stuck at present -

  1. How to generate style mixing using our test sample of 2 person rather than seeds?
  1. While generating fake or stylemix as per repository we generally give seeds values as input.Like seed=5.
    1. What is range of seed values to be used? How to determine that?
    2. In style mix we give row-seed and col-seed, but each seed will generate random image. How can we take control over this image generation, suppose can we choose to stylemix only male person and not random male/female?

Thanks in Advance!

generative-adversarial-network
1 Answers

I would say it is impossible to use 2 real images, and then compute the intermediate states between them. The reason is as follows. The first image is generated from a random vector (e.g. a vector from a normal distribution). The second image is also generated from a second random vector. Now, an intermediate image (e.g. the "mixed" image) is computed from the vector in between those previous 2 vectors. For a real image, you don't know what random vector can create this real image, and so you don't have the intermediate vectors to create "mixed" images.

user_72393470
Related