HyperAI超神经

图像生成

图像生成(合成)是从现有数据集中生成新图像的任务。无条件生成指从数据集中无条件地生成样本,即$p(y)$;而有条件生成则是基于标签从数据集中有条件地生成样本,即$p(y|x)$。本节展示无条件生成的最新技术排行榜,其他类型图像生成请参见子任务。图像生成在计算机视觉中具有重要应用价值,能够用于数据增强、艺术创作和虚拟现实等领域。

ImageNet 256x256
Discriminator Guidance
CIFAR-10
GMem
ImageNet 64x64
CTM (NFE 1)
FFHQ 256 x 256
Anycost GAN
ImageNet 512x512
MAGVIT-v2
CelebA 64x64
HDCGAN
ImageNet 32x32
StyleGAN-XL
LSUN Bedroom 256 x 256
StyleGAN (DINOv2)
STL-10
LSUN Churches 256 x 256
BOSS
ImageNet 128x128
ADM-G
FFHQ 1024 x 1024
Efficient-VDVAE
CelebA-HQ 256x256
BOSS
CelebA 256x256
StyleSwin
MNIST
Locally Masked PixelCNN (8 orders)
FFHQ-U
Alias-Free-R
FFHQ
Anycost GAN
Binarized MNIST
CR-NVAE
CelebA-HQ 1024x1024
StyleSwin
CIFAR-100
LeCAM (StyleGAN2 + ADA)
AFHQ Cat
Vision-aided GAN
LSUN Cat 256 x 256
Projected GAN
AFHQV2
Polarity-StyleGAN3
CelebA-HQ 128x128
COCO-GAN
Fashion-MNIST
GLF+perceptual loss (ours)
TextAtlasEval
AFHQ Dog
Vision-aided GAN
Cityscapes
GANformer
CLEVR
Projected GAN
LSUN Horse 256 x 256
StyleGAN2
AFHQ Wild
Vision-aided GAN
CelebA 128x128
U-Net GAN
Places50
SinDiffusion
ARKitScenes
GAUDI
CUB 128 x 128
Projected GAN
Pokemon 256x256
StyleGAN-XL
Replica
Stanford Cars
Projected GANs
Stanford Dogs
Projected GAN
VizDoom
VLN-CE
ADE-Indoor
CAT 256x256
StyleGAN2 + DA + RLC (Ours)
CelebA-HQ 64x64
COCO-GAN
CIFAR-10 (10% data)
DiffAugment-StyleGAN2
CIFAR-10 (20% data)
DiffAugment-StyleGAN2
FFHQ 128 x 128
Anycost GAN
FFHQ 512 x 512
Anycost GAN
LSUN Bedroom
StyleGAN
LSUN Bedroom 64 x 64
WGAN-GP + TTUR + Alex-Adam
MetFaces
MetFaces-U
ObjectsRoom
Pokemon 1024x1024
StyleGAN-XL
ShapeStacks
Stacked MNIST
VAEBM
AFHQ-v2 64x64
FFHQ 64x64
SiDA-EDM
iNaturalist 2019
StyeGAN2 + NoisyTwins
LSUN Bedroom 128 x 128
LadaGAN
LSUN Car 512 x 384
Polarity-StyleGAN2
Oxford 102 Flowers 256 x 256
MSG-StyleGAN
RC-49
cDR-RS
25% ImageNet 128x128
LeCAM + DA
CelebA
CelebA-HQ
DDPM
CelebA-HQ 512x512
WaveDiff
Cityscapes-25K 256x512
SB-GAN
Cityscapes-5K 256x512
SB-GAN
EMNIST-Letters
Spiking-Diffusion
FFHQ 64x64 - 4x upscaling
PFGM++
GQN
ImageNet 256x256 - 1 labeled data per class
ImageNet 256x256 - 1% labeled data
DPT
ImageNet 256x256 - 2 labeled data per class
ImageNet 256x256 - 5 labeled data per class
Indian Celebs 256 x 256
MSG-StyleGAN
KMNIST
Landscapes 256 x 256
CIPS
LLVIP
pix2pix
LSUN
BigGAN + gSR
LSUN Car 256 x 256
StyleGAN2
LSUN tower 64x64
DDPM-IP
Multi-dSprites
GENESIS
NASA Perseverance
Oxford 102 Flowers 128x128
QSNGAN
Satellite-Buildings 256 x 256
CIPS
Satellite-Landscapes 256 x 256
CIPS
SDSS Galaxies