Generative Models Quiz

Which of the following correctly describes the role of the encoder and decoder in a Variational Autoencoder (VAE) for image generation?

• The encoder compresses the image into a fixed representation, and the decoder reconstructs the exact same image

• The encoder maps the image to a distribution in latent space, and the decoder samples from this distribution to generate similar images

• The encoder identifies key features in the image, and the decoder amplifies these features

• The encoder removes noise from the image, and the decoder enhances image quality

Which statement(s) correctly distinguish a Variational Autoencoder (VAE) from a standard Autoencoder (AE)?

• VAEs produce deterministic encodings, while AEs produce probabilistic latent representations.

• VAEs learn a continuous latent space by sampling from a distribution, whereas AEs learn a direct mapping without probabilistic sampling.

• VAEs cannot reconstruct inputs accurately due to random noise in sampling, while AEs always perform perfect reconstruction.

• AEs use an encoder–decoder structure, while VAEs do not.

Which of the following are true about Diffusion Models used in image generation?

• They learn to reverse a noising process applied to training images.

• They require adversarial training with a discriminator network.

• They can produce high-quality images by iteratively denoising samples from random noise.

• They do not use latent variables at all.

Which of the following statements about latent space is/are correct?

• A well-trained latent space allows smooth interpolation between data samples.

• Latent dimensions always correspond directly to interpretable features (e.g., “hair color,” “background color”).

• A “latent vector” can be sampled from a prior (e.g., Gaussian) to generate new outputs.

• High-dimensional latent spaces automatically guarantee high-quality generation.

In large language models (LLMs) such as GPT-3.5, which techniques are commonly used to improve training and generation quality?

• Masked language modeling (like in BERT)

• Teacher forcing for sequence-to-sequence training

• Next-token prediction on large-scale unlabelled text corpora

• Reinforcement learning with human feedback (RLHF)

Which of the following is a challenge specific to text-to-video generation compared to text-to-image?

• Synchronizing multiple frames to maintain temporal consistency

• Generating color images

• Handling variable-length text prompts

• Converting text to embeddings

You want to ensemble different generative models (e.g., a diffusion model and a VAE) to leverage their strengths. Which approach(s) correctly describe possible ensembling strategies?

• Train both models on the same data and then average their generated outputs pixel by pixel.

• Use one model’s latent representation as input to another model (e.g., VAE → Diffusion) to refine generation.

• Randomly pick outputs from either the diffusion model or the VAE.

• Use a gating network that decides which model to use based on input constraints or quality measures.

When implementing a text-to-image system for an e-commerce platform that needs to generate product variations with specific attributes (color, style, etc.), which approach would be most suitable?

• Fine-tuning a pre-trained DALL-E model on your product catalog

• Using a GAN architecture trained from scratch on your products

• Implementing ControlNet with a Stable Diffusion base model

• Using a series of image-to-image translations with cyclical consistency

What is the primary difference between latent diffusion models (like Stable Diffusion) and GANs for image generation?

• Latent diffusion models use deterministic sampling while GANs use probabilistic sampling

• Latent diffusion models generate images by progressively removing noise, while GANs use adversarial training

• Latent diffusion models operate in pixel space, while GANs operate in latent space

• Latent diffusion models require more training data than GANs

When implementing a multilingual chatbot that needs to understand and generate responses in multiple languages, which technique would provide the best balance of performance and efficiency?

• Training separate language models for each supported language

• Using a large multilingual transformer model with language-specific adapters

• Implementing machine translation before and after a monolingual model

• Fine-tuning a multilingual model on conversation data across all target languages