Skip to main content

img2img

  • img2img
    • changing a given image in a specific or controlled way
  • Pix2Pix
    • 成对训练数据
    • 训练快
    • 数据集
      • 需要一定程度的准确性
  • Pix2PixHD - 2017
    • 2048x1024
    • photorealistic image-to-image
    • semantic label maps -> photo-realistic images
    • synthesizing portraits from face label maps
    • conditional GANs
  • Vid2Vid - 2018
  • SPADE (Spatially-Adaptive Normalization) - 2019
  • Pix2PixSC (Style-Consistent) (2020)
  • TransGAN (2021)
  • Pix2Pix-Zero (2023)
  • CycleGAN
    • 无需成对训练数据
    • 训练慢
  • Pix2Pix with Temporal Consistency
  • RecycleGAN - CycleGAN
  • FOMM (First Order Motion Model)
  • StyleGAN2 with temporal consistency
  • Transformer-based Video Inpainting
  • few-shot learning
  • perceptual loss 感知损失
    • 改善生成质量
  • total variation loss 总变差损失
    • 保持图像平滑
    • 处理模糊和噪声
  • Style Loss 风格损失
    • 风格迁移
  • Adversarial Loss 对抗损失
    • 生成器和判别器
  • U-Net