search

danielroich / PTI

Official Implementation for "Pivotal Tuning for Latent-based editing of Real Images" (ACM TOG 2022) https://arxiv.org/abs/2106.05744
MIT License
905 stars 114 forks source link

generate random faces #10

Closed loboere closed 3 years ago

loboere commented 3 years ago

amazing work. I would like to know if it is possible to generate random faces with a seed like nvidia's stylegan. Try to do this, but the generated messages are full of artifacts. new_G.synthesis(torch.from_numpy(np.random.rand(1,18,512)).float().to("cuda"),noise_mode='const')

danielroich commented 3 years ago

Hi @loboere. Yes, it is possible.

You missed the mapping network. The synthesis network does not accept the z vector sampled from the normal distribution. this vector is first passed to a mapping network which outputs a vector in the W space, and then this vector is passed to the synthesis network.

The code you are interested in is: z = torch.randn([1, new_G.z_dim]).cuda() w = new_G.mapping(z, None, truncation_psi=0.5, truncation_cutoff=8) image = new_G.synthesis(w, noise_mode='const', force_fp32 = True) plot_syn_images([image])

When the new_G is the output generator from PTI.

image

Hope this helps, Daniel

loboere commented 3 years ago

works! Thank you