Викторина по стабильной диффузии

Ответы и пояснения появляются по мере прохождения теста.

1. На высоком уровне, как диффузионная модель генерирует изображение?
   • Он начинается со случайного шума и неоднократно удаляет шум, чтобы сформировать изображение.
   • Он копирует наиболее подходящее изображение из своего обучающего набора.
   • Он рисует фигуры одним штрихом за раз, как человек.
   • Он объединяет фотографии, найденные в Интернете.
2. Where does the original Stable Diffusion v1 pipeline perform its learned denoising work?
   • Only in an unchanged image filename
   • Only in a single scalar with no spatial features
   • Without any learned neural model
   • In an autoencoder latent representation rather than directly over full-resolution pixels
3. How does text condition the documented original Stable Diffusion v1 denoiser?
   • The prompt only names the output file
   • Each prompt word must be printed literally into every output
   • CLIP text embeddings provide conditioning through cross-attention
   • The model ignores all text information
4. What does classifier-free guidance combine in the documented sampling formula?
   • Two unrelated image filenames
   • Conditional and unconditional denoising predictions
   • Only the red and blue display channels
   • A guarantee that every increased scale improves quality
5. Which statement accurately qualifies a change to sampler step count?
   • Every additional step must improve every model’s image
   • Step count never changes any computation
   • A single step guarantees perfect image fidelity
   • Quality and elapsed time depend on the model, sampler and settings; more steps are not universally better
6. What does the original Stable Diffusion project’s publication of model artifacts enable?
   • Inspection and supported extension under the stated model and license conditions
   • Unrestricted use of every model without any conditions
   • A guarantee that all devices can run every checkpoint
   • Proof that every possible fine-tune is compatible
7. What additional input type does the cited ControlNet method provide for diffusion generation?
   • Only the final output filename
   • Spatial conditions such as edges, pose or depth
   • A mandatory list of every training image
   • Only an encryption password
8. With the documented factor-eight spatial downsampling, what latent height and width correspond to a 512-by-512 v1 image?
   • 512 by 512
   • 128 by 128
   • 64 by 64
   • 8 by 8