引导式测验 · 中等 等级

LaMa 分辨率 - 强大的修复测验

了解 LaMa 如何结合傅立叶卷积、高感受野损失和大掩模进行图像修复及其报告的限制。

相关引导路径喇嘛决议坚固修复
问题 1 的 8

赋予 LaMa 图像全球视野的标志性建筑组件是什么？

预览每个问题8 问题

当您参加测验时，就会出现答案和解释。

  1. 赋予 LaMa 图像全球视野的标志性建筑组件是什么？
    • 快速傅立叶卷积
    • Transformer 自注意力块
    • 循环 LSTM 层
    • 胶囊网络
  2. 为什么 LaMa 比许多早期的修复模型更能处理大型面具？
    • 它只适用于小图像
    • 它的全球感受野让它能够跨越大洞继续结构
    • 它从边缘随机复制像素
    • 它拒绝填充超过固定尺寸的掩模
  3. What resolution behavior did the LaMa paper report?
    • It only works at one exact training resolution
    • It works only below the training resolution
    • It generalized to some resolutions higher than those used during training in the evaluated settings
    • It accepts no image input
  4. 由于其频域处理，LaMa 能够特别出色地处理哪种纹理？
    • 纯随机噪声
    • 只有纯色
    • 仅限文本字符
    • 重复、周期性的纹理，如砖块或瓷砖
  5. Compared with large diffusion-based inpainting systems, how is LaMa commonly positioned?
    • As a comparatively efficient convolutional inpainting method
    • As a much slower model that requires a data center
    • As a video-only model
    • As a tool that never needs an image
  6. 哪种损失有助于 LaMa 生成逼真的填充而不是模糊的平均值？
    • 一个简单的像素均方误差损失
    • 对抗性损失与知觉损失相结合
    • 根本没有损失函数
    • 仅测量文件大小的损失
  7. What additional loss component does the LaMa paper emphasize for large-mask inpainting?
    • 语音识别损失
    • A bitrate-only loss
    • A perceptual loss with a high receptive field
    • A loss that ignores image structure
  8. Who introduced LaMa in the 2021 paper?
    • Jonathan Frankle and Michael Carbin
    • Cees Taal and colleagues
    • Timothy Dozat and Christopher Manning
    • Roman Suvorov and colleagues