Scholarlink

This chapter introduces 3D image and video generation using GANs. For 3D image generation, frameworks like Visual Object Networks decompose tasks into shape, projection, and texture modules. PrGAN infers 3D shapes from 2D views via adversarial training. Video generation models include Video-GAN (separating static/dynamic components) and MoCoGAN (decoupling content/motion spaces). MoCoGAN-HD extends StyleGAN for high-resolution video synthesis by manipulating latent vectors. MD-GAN uses a two-stage pipeline (content generation + motion refinement) with Gram matrix constraints. The chapter addresses challenges in temporal coherence, resolution, and dataset scarcity (e.g., ShapeNet for 3D shapes).

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

3D Image and Video Generation

  • 章节

  • 发表日期：

    2026年

查看全文
  • Peng Long,
  • Xiaozhou Guo

摘要

This chapter introduces 3D image and video generation using GANs. For 3D image generation, frameworks like Visual Object Networks decompose tasks into shape, projection, and texture modules. PrGAN infers 3D shapes from 2D views via adversarial training. Video generation models include Video-GAN (separating static/dynamic components) and MoCoGAN (decoupling content/motion spaces). MoCoGAN-HD extends StyleGAN for high-resolution video synthesis by manipulating latent vectors. MD-GAN uses a two-stage pipeline (content generation + motion refinement) with Gram matrix constraints. The chapter addresses challenges in temporal coherence, resolution, and dataset scarcity (e.g., ShapeNet for 3D shapes).