ControlNet 测验

当您参加测验时，就会出现答案和解释。

1. What does ControlNet add to a pretrained text-to-image diffusion model?
   • A spatial conditioning path that can guide structure using inputs such as edges or pose maps
   • A guarantee that every generated pixel matches the input map exactly
   • A replacement for the text encoder and denoiser
   • A tool that predicts the original training image
2. ControlNet 中“零卷积”层的作用是什么？
   • 他们删除了原始模型权重
   • 他们从零权重开始，因此训练从原始模型的行为开始
   • 他们将图像转换为灰度
   • 他们在提示中添加随机噪音
3. 其中哪一个是有效的 ControlNet 调节输入？
   • 数字电子表格
   • 音频波形
   • OpenPose 骨架或深度图
   • 数据库架构
4. ControlNet 与稳定扩散等基本扩散模型有何关系？
   • 它完全取代了基本模型
   • 它只能在没有任何基础模型的情况下工作
   • 它通过互联网在单独的计算机上运行
   • 它作为可训练分支连接，而基本模型保持冻结状态
5. In a ControlNet workflow, what roles can text and spatial conditioning play?
   • Text can describe semantic/style content while a spatial map supplies structural guidance
   • The spatial map proves the text prompt is factually true
   • Text always supplies exact pixel positions and the map supplies the object name
   • The text prompt is ignored whenever a map is present
6. What training-data scale does the ControlNet paper report as robust for training its controls?
   • Only datasets with exactly 50 examples
   • Both small datasets under 50,000 examples and large datasets over 1 million examples
   • Only datasets larger than 100 million examples
   • No training data is used
7. What does the ControlNet paper report about combining conditioning controls?
   • It permits only one edge map and requires no prompt in every case
   • It accepts text prompts but no spatial controls
   • It tests single or multiple conditions, with or without a text prompt
   • It requires exactly two depth maps for each image
8. Which pretrained part does ControlNet reuse as a strong backbone for learning spatial controls?
   • A database of user prompts
   • A language-model tokenizer only
   • A separately trained classifier for every edge map
   • The deep encoding layers of a pretrained diffusion model