Layout

Layout Article

• [ICCV19] LayoutVAE: Stochastic Scene Layout Generation From a Label Set
  • model: VAE
  • task: 根据 label set, per label layouts in existing image 进行有条件生成
• [CVPR21] LayoutTransformer: Layout generation and completion with self attention
  • model: transformer
  • task: completion
• [MM21] LayoutGAN++: Constrained Graphic Layout Generation via Latent Optimization
  • model: GAN
• [ECCV22] BLT: Bidirectional layout transformer for controllable layout generation
  • model: transformer & BERT
  • task: 根据 element, category 进行有条件生成，无条件生成
  • pros:
  • hierarchical mask sampling policy，一个改进了的 mask 方法，而不是直接应用 Transformer 的 mask
  • cons:
  • 不能处理元素之间的关系
• [CVPR23] LayoutFormer++: Conditional Graphic Layout Generation via Constraint Serialization and Decoding Space Restriction
  • model: transformer
  • condition: 根据 type, relationship, type & size 进行有条件生成, refinement, completion, 无条件生成
  • pros:
  • 灵活性和可控性强
  • A constraint serialization scheme which can handle diverse constrain
  • A sequence to sequence layout generation method
• [CVPR23] LayoutDM: Discrete Diffusion Model for Controllable Layout Generation
  • model: diffusion
  • task: Category \(\to\) Size+Position, Category+Size \(\to\) Position, Completion, 无条件生成
  • pro:
  • padding approach to model highly structured layout data
  • masking and logit adjustment during the inference
  • cons:
  • 作者在进行训练和评估模型时，剔除了数据集中元素个数 \(\ge50\) 的 layout 数据，在输入文档具有更多元素时，模型的性能会降低
• [CVPR23] LDGM: Unifying Layout Generation with a Decoupled Diffusion Model
  • model: diffusion
  • task: 根据 \((c,x,y,w,h)\) 五种属性进行有条件生成, refinement, completion, 无条件生成
  • pros:
  • 对于类别属性 \((c)\)、位置属性 \((x, y)\) 和大小属性 \((w, h)\) 构造了三条 timeline 分别添加噪声

---

Last update: August 29, 2023
Created: June 14, 2023