Papers of the day   All papers

Analyzing and Improving the Image Quality of StyleGAN


Xander Steenbrugge: Whoa, StyleGANv2 is out! - Significantly better samples (better FID scores & reduced artifacts) - No more progressive growing - Improved Style-mixing - Smoother interpolations (extra regularization) - Faster training Paper: Github:

6 replies, 611 likes

roadrunner01: Analyzing and Improving the Image Quality of StyleGAN pdf: abs: github:

5 replies, 295 likes

Ming-Yu Liu: StyleGAN2 is out.

1 replies, 213 likes

Alexia Jolicoeur-Martineau: NVIDIA basically fixed most of the issues with their architecture. They now use gradient penalty only 1/16 of the times making it much faster and they replaced progressive growing by a modified MSG-GAN (@AnimeshKarnewar). The amount of work done is insane for one single paper.

3 replies, 173 likes

Jonathan Fly ๐Ÿ‘พ: StyleGAN 2: Mostly We Got Rid of The Blobs Also: No 'phase artifacts' (I hadn't noticed them but obvious in retrospect), easier-to-navigate latent spaces mean better encoding, and ~25% more efficient. pdf: code (coming-later):

2 replies, 137 likes

Mikael H Christensen: #StyleGAN2 is out and looking great! I have created a Google Colab notebook to make it easier to get started:

2 replies, 130 likes

Ben Nimmo: Perhaps counter-intuitively, backgrounds are even harder, because there's more variation in scenery than there is in faces.

2 replies, 82 likes

Jaakko Lehtinen: This just in: we have updated the #StyleGAN2 paper with estimates of total GPU time and energy consumption, with a breakdown over the different facets of the research project. See new Appendix F and Table 5.

1 replies, 65 likes

Jaakko Lehtinen: #StyleGAN2 out now

1 replies, 57 likes

ๅฐ็Œซ้Šใ‚Šใ‚‡ใ†๏ผˆใŸใ‹ใซใ‚ƒใ—ใƒปใ‚Šใ‚‡ใ†๏ผ‰: ่ซ–ๆ–‡ Analyzing and Improving the Image Quality of StyleGAN StyleGAN2 โ€” Official TensorFlow Implementation

1 replies, 43 likes

Drew Harwell: These are all computer-generated faces. (From new Nvidia AI research:

1 replies, 37 likes

Miles Brundage: Interesting: "It turns out that our improvements to StyleGAN make it easier to detect generated images using projection-based methods, even though the quality of generated images is higher."

1 replies, 35 likes

Jane Lytvynenko ๐Ÿคฆ๐Ÿฝโ€โ™€๏ธ๐Ÿคฆ๐Ÿฝโ€โ™€๏ธ๐Ÿคฆ๐Ÿฝโ€โ™€๏ธ: this tech has gotten so much better in such a short span of time ๐Ÿ˜ฌ

5 replies, 32 likes

Giorgio Patrini ๐Ÿ›ก๏ธ๐Ÿ‘พ: This analysis is excellent @benimmo but please careful: some of these telltales are already solved by StyleGAN v2, and won't be there at the next disinfo automated campaign exploiting synthetic photos

1 replies, 21 likes

Kyle McDonald: NVIDIA just released the latest tech for image generation. I can still find a few artifacts if i zoom in at 1024x1024, but i would guess weโ€™ve passed human recognition at 512x512

2 replies, 19 likes

Koichi Hamada: StyleGAN2: "Analyzing and Improving the Image Quality of StyleGAN" - Github: - ArXiv: - Video:

0 replies, 9 likes

Jeremy Cowles: Update to StyleGAN focused on analyzing the previous architecture, I'm super curious to see what they found:

1 replies, 8 likes

Emmett Macfarlane: Whoa...

0 replies, 7 likes

jess: @LargeCardinal On a similar note, Nvidia released StyleGAN2 a few months back. The quality and accessibility of these images is second to none, especially with sites such as paper's worth reading: also have a git repo:

1 replies, 6 likes

Shanthi Kalathil: With even just a tiny bit of imagination, it becomes clear that current efforts to develop widespread digital literacy and strengthen democratic resilience are likely to be quickly outpaced.

0 replies, 6 likes

Andy Baio: More info in this thread if you want to learn more.

1 replies, 5 likes

Daisuke Okanohara: StyleGAN2:1) Replace Instance Norm with Weight Norm to remove droplet artifact 2) Introduce path length regularization to minimize perceptual path length 3) skip-connected generator, residual discriminator, no progressive growing.

0 replies, 5 likes

Aaron Gokaslan: StyleGAN v2 is out!

0 replies, 3 likes

Mรกlaga Artificial Intelligence: StyleGANv2. The authors did a big homework and fixed "light spots" and "burnt skin" artifacts. Also, they proposed an improved method for finding real images in latent space (but still via backprop). ๐Ÿ”Ž ๐Ÿ“ ๐Ÿ“‰ @loss_function_porn

0 replies, 2 likes

Roelof Pieters: StyleGan is back with a V2 overhaul! - Much better quality and FID scores - Removing artefacts by demodulation - No progressive growing - New MSG-GAN architecture - Quicker training Paper: Code:

0 replies, 2 likes

Ruqiya: ุชุญุณูŠู† ู„ู„ styleGAN ูˆุฑู‚ุฉ ู†ุดุฑุช ุจู‡ุฐุง ุงู„ุดู‡ุฑ 3 ุฏูŠุณู…ุจุฑ 2019 ุนู†ูˆุงู†ู‡ุง Analyzing and Improving the Image Quality of StyleGAN ุนู„ู‰ ุงู„ุฑุงุจุท ุญุงู„ูŠุง ู…ู‡ุชู…ุฉ ุจุงู„ GANs

0 replies, 1 likes

Parand Darugar: StyleGAN, the system that creates those eerily realistic portraits you saw at , has been updated to be even more effective: Video: Paper: So many cool things coming out of NeurIPS, wish I was there.

0 replies, 1 likes

Deeptrace: @GlennF @razhael Yet it is important to point out that those signs may soon disappear. Version 2 of the StyleGAN model was open sourced just days ago, and claimed to have solved some of the known issues in generation.

1 replies, 0 likes


Found on Dec 12 2019 at

PDF content of a computer science paper: Analyzing and Improving the Image Quality of StyleGAN