less than 1 minute read

Meta info.

TL; DR

ํ•™์Šต์ด ๋ถˆ์•ˆ์ •ํ•œ GAN์˜ 1) Loss ์ˆ˜์ • 2) ์ตœ์‹  architecture ์ ์šฉํ•˜์—ฌ SOTA

image.png

image.png

image.png

The GAN is dead; long live the GAN! A Modern GAN Baseline (NeurIPS2024)

  • TL; DR: ํ•™์Šต์ด ๋ถˆ์•ˆ์ •ํ•˜๋‹ค๋Š” GAN์˜ 1) Loss ์ˆ˜์ • 2) ์ตœ์‹  architecture ์ ์šฉํ•˜์—ฌ SOTA

Problem States

GAN ํ•™์Šต์˜ ๋ฌธ์ œ์ ์œผ๋กœ mode dropping, non-convergence์ด ์ž˜ ํ•ด๊ฒฐ๋˜์ง€ ์•Š์Œ

  • ๊ฒฝํ—˜์ ์ธ ๋ฐฉ์‹์— ์˜์กดํ•˜๊ฑฐ๋‚˜, GAN ๊ธฐ๋ฐ˜ ๋ฐฉ๋ฒ•๋ก ๋“ค์ด ์ง€๋‚˜์น˜๊ฒŒ ๋ณต์žกํ•จ

Suggestions

์ตœ์‹  ๋”ฅ๋Ÿฌ๋‹ ๋ฐฉ๋ฒ•๋ก ์„ ์ ์šฉํ•˜์—ฌ ๊ธฐ์กด ๋ฌธ์ œ์ ์„ ์ƒ๋‹น์ˆ˜ ๊ฐœ์„ ํ•œ GAN Baseline R3GAN ์ œ์•ˆ

- Loss ์ˆ˜์ •: RpGAN
    - ์ˆ˜ํ•™์ ์œผ๋กœ local convergence ๋ณด์žฅ ์ฆ๋ช…(appendix) = ๊ทผ๋ณธ์ ์ธ ๊ธฐ์กด ๋ฌธ์ œ์  ํ•ด๊ฒฐ
    - ๊ทผ๋ณธ GAN architecture ํšŒ๊ท€ = ์—ฌํƒ€ ๊ฒฝํ—˜์ ์ธ trick ๋Œ€์ฒด
- ์ตœ์‹ ์˜ minimalํ•œ architecture ์ ์šฉ: ResNet, U-Net, ViT ๋“ฑ - Results
- FFHQ, ImageNet, CIFAR, Stacked MNIST ๋“ฑ ๋‹ค์–‘ํ•œ ๋ฐ์ดํ„ฐ์…‹์—์„œ StyleGAN2 ๋น„๋กฏ, ์ตœ์‹  GAN/Diffusion model๊ตฐ ์ œ์น˜๊ณ  SOTA

Personal note. ์ €์ž๋“ค์ด ๊ฐœ์„ ํ•œ ๊ฒƒ์ด ๋”ฑ ๋‘๊ฐ€์ง€๋กœ ๊ฐ„์†Œํ•œ๋ฐ ํšจ๊ณผ๊ฐ€ ๋šœ๋ ทํ•ฉ๋‹ˆ๋‹ค. StyleGAN2๋ž‘ ๋น„๊ตํ•˜๋Š”๋ฐ, ๋ณต์žกํ•˜๋‹ค๋Š” ์ •๋„๋งŒ ์ดํ•ดํ•˜๊ณ  ์žˆ์–ด๋„ ์ˆ˜์›”ํ• ๋“ฏ.