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
  • 引入时序判别器来保持时间一致性
  • https://arxiv.org/abs/1808.06601
  • NVlabs/few-shot-vid2vid - 2019
    • https://arxiv.org/abs/1910.12713
    • -> NVlabs/imaginaire
• 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

---

• img2img-turbo
  • One-step image-to-image with Stable Diffusion turbo
• junyanz/pytorch-CycleGAN-and-pix2pix
  • Image-to-Image Translation in PyTorch
• taesungp/contrastive-unpaired-translation
  • 2020
  • https://arxiv.org/abs/2007.15651
• https://machinelearningmastery.com/a-gentle-introduction-to-pix2pix-generative-adversarial-network/