Unit 6: Generative Vision Models - Practice Quiz

1 What is the main purpose of the latent space in a variational autoencoder?

2 What does a variational autoencoder typically use to generate a new sample?

3 Which two networks form the basic architecture of a GAN?

4 What is the role of the discriminator in a GAN?

5 What does the generator in a GAN produce?

6 What type of layers are commonly used in a DCGAN for image processing?

7 What is a common function of transposed convolutional layers in a DCGAN generator?

8 What type of task is CycleGAN mainly designed to perform?

9 What does cycle consistency encourage in CycleGAN?

10 What is a key feature of StyleGAN?

11 What is image generation?

12 What is the goal of image-to-image translation?

13 What does image super-resolution attempt to do?

14 What is usually the input to a single-image super-resolution model?

15 What does CLIP learn to align?

16 How can CLIP be used for zero-shot image classification?

17 What is the main goal of interpretability techniques in deep learning?

18 What does a Grad-CAM heatmap show?

19 What does a saliency map usually indicate?

20 Which visual pattern often represents an important region in a saliency map?

21 A variational autoencoder represents an input using and . Which expression allows latent samples to be generated while preserving gradient-based training?

22 A VAE produces realistic reconstructions, but latent samples are poorly organized and generation from the prior fails. Which training change most directly addresses this issue?

23 During GAN training, the discriminator correctly rejects nearly every generated image, and the generator receives very small gradients under the original minimax loss. Which generator objective is commonly used to improve its gradient signal?

24 A trained GAN generates sharp images, but many different latent vectors produce nearly identical faces. What problem does this behavior indicate?

25 Which generator design is most consistent with the standard DCGAN architecture?

26 A DCGAN generator creates visible checkerboard patterns in otherwise plausible images. Which modification is most likely to reduce these artifacts?

27 A CycleGAN is trained to translate photographs into paintings without paired examples. What does cycle-consistency loss require for a photograph ?

28 When translating horses to zebras, a CycleGAN unnecessarily changes the color of the sky. Which added objective can encourage already suitable target-domain content to remain unchanged?

29 What is the main purpose of StyleGAN's mapping network that transforms into an intermediate latent code ?

30 In StyleGAN, changing the learned per-layer noise inputs while keeping the style code fixed primarily changes which image properties?

31 Two image generators are evaluated using Fréchet Inception Distance (FID). Model A has an FID of 18, while Model B has an FID of 42 on the same dataset and feature extractor. What is the best interpretation?

32 Linear interpolation between two nearby latent vectors produces abrupt, unrealistic intermediate images. What does this most strongly suggest about the learned latent space?

33 A researcher has aligned pairs of building facade labels and corresponding photographs. Which training approach is most appropriate for learning the translation?

34 In a paired image-to-image translation model, increasing the weight of an reconstruction loss excessively is most likely to produce which result?

35 A super-resolution system trained only with mean squared error achieves high PSNR but produces overly smooth textures. Which additional loss is most suitable for improving perceptual detail?

36 A super-resolution model receives an image of size . Assuming both spatial dimensions are scaled by four, what output size should it produce?

37 To perform zero-shot image classification with CLIP, how should the predicted class be selected?

38 In CLIP's contrastive training, what should happen to the similarity of a correctly matched image-caption pair relative to mismatched pairs in the same batch?

39 An engineer repeatedly masks small regions of an image and records the resulting decrease in a classifier's confidence. Which interpretability technique is being applied?

40 Why is the last convolutional layer commonly used to construct a Grad-CAM visualization?

41 A VAE trained with objective uses a highly expressive autoregressive decoder. During training, the KL term converges to nearly zero and reconstructions become independent of . Which intervention most directly addresses this failure while preserving the VAE formulation?

42 Let . Which sampling expression permits low-variance pathwise gradients of the reconstruction term with respect to ?

43 Early in GAN training, a discriminator confidently rejects generated samples, so . Why is the non-saturating generator loss usually preferred to minimizing ?

44 A Wasserstein GAN critic is optimized without a sigmoid output. Which property is essential for the critic objective to correspond to the Kantorovich-Rubinstein dual of the Wasserstein-1 distance?

45 A DCGAN generator produces periodic checkerboard artifacts after each upsampling stage. Which architectural change most directly reduces artifacts caused by uneven overlap in transposed convolutions?

46 A CycleGAN translates photographs to maps and back with low adversarial and cycle-consistency losses, yet the generated maps contain imperceptible high-frequency signals encoding the source photograph. What does this reveal about the objective?

47 For unpaired domains and , suppose maps every input to the same realistic image in . Which CycleGAN term most directly penalizes this many-to-one collapse?

48 In a StyleGAN synthesis layer, modulation scales each input feature channel according to a style vector, and demodulation then normalizes each output channel's effective weights. What is the main purpose of demodulation?

49 A StyleGAN model injects independent learned-strength noise into synthesis layers. Altering noise at fine-resolution layers while holding styles fixed should primarily change which image property?

50 In classifier-free guidance for a diffusion image generator, the guided noise estimate is . What is the most likely effect of choosing a very large ?

51 In pix2pix, replacing a full-image discriminator with a PatchGAN discriminator changes the learned adversarial criterion. Which behavior follows most directly?

52 A conditional image-to-image task has several valid outputs for each input, but training uses only a deterministic generator with a strong reconstruction loss. What failure is most expected?

53 Two super-resolution systems are evaluated on the same test set. Model P obtains higher PSNR but looks smoother; model G obtains lower PSNR but more realistic textures. Which explanation best accounts for this result?

54 A super-resolution network is trained to invert bicubic downsampling but is deployed on images degraded by unknown blur, sensor noise, and JPEG compression. Why can performance collapse despite excellent benchmark PSNR?

55 CLIP is trained on a batch of matched image-text pairs using a symmetric contrastive loss over an similarity matrix. If two different captions in the batch accurately describe the same visual content but are not paired with the same image, how does the standard loss treat them?

56 A zero-shot CLIP classifier performs poorly when class names are embedded as isolated words but improves when prompts such as "a photo of a {class}" are averaged across templates. What is the best explanation?

57 A saliency method produces nearly unchanged explanations after the model's learned weights are randomly reinitialized, although its maps still look edge-like. What is the strongest conclusion from this sanity check?

58 For class score and convolutional feature maps , standard Grad-CAM computes . Which expression then forms the class activation map?

59 Grad-CAM is computed from the final convolutional layer of a classifier and produces a coarse heatmap that misses small diagnostic structures. Which change most directly improves spatial resolution, although it may reduce semantic specificity?

60 Integrated Gradients attributes features along the straight-line path from baseline to input . Assuming differentiability, which property distinguishes it from a raw input-gradient saliency map?