1What did “The Painting Fool” contribute to AI art?
It spawned the movement for computer art
2What is Neural Style Transfer?
A technique that transfers the style of one image to another using convolutional neural networks
3What is an additional use of Neural Style Transfer?
Bringing back colors in black-and-white images
4What are CNNs especially good at?
Abstract feature extraction
5What is the main drawback of CNNs?
They require long training times with large datasets
6What are the two parts of a GAN?
The generator (creates images) and the discriminator (judges images as real or fake)
7How does the discriminator help the generator in GANs?
By sending feedback (backpropagation) so the generator can learn from mistakes
8What is the competition in GANs often described as?
A battle between a faker (generator) and a detector (discriminator)
9What is vanishing gradients in GAN training?
When the discriminator is too good, preventing the generator from learning
10What is mode collapse in GANs?
When the generator produces limited variations of outputs instead of diverse images
11What is non-convergence in GANs?
When training fails to stabilize and both models fail to improve
12What is alignDRAW?
A recurrent neural network model that generates images from text labels
13What was unique about alignDRAW compared to earlier networks?
It generated images based on labels instead of just recognizing labels from images
14What was the AI Spring in image generation?
The rise of transformer-based model architectures
15What is CLIP?
Contrastive Language-Image Pre-training, a technique to train models for image and text understanding using a shared embedding space
16How many image-text pairs were used in CLIP pre-training?
400 million
17What type of architecture does CLIP use?
A dual-encoder with one encoder for images (Vision Transformer/ResNet) and one for text (transformer-based model)
18What is the goal of contrastive loss in CLIP?
To increase similarity of correct image-text pairs and decrease similarity of mismatched pairs
19What does CLIP enable after training?
Zero-shot prediction for unseen images without fine-tuning
20How does CLIP perform classification of unseen images?
By comparing cosine similarity between image embeddings and text embeddings
21What is diffusion in image generation?
A process that converts noise into images by gradually denoising step by step
22How does a diffusion model learn?
It predicts and removes noise from noisy training images until a clean image is formed
23What is a noise schedule in diffusion models?
The rate at which noise is removed during the denoising process
24What happens if the number of diffusion steps is increased?
The generated image becomes higher quality and more detailed
25What is the role of the encoder in diffusion-based models?
To compress an image into a latent space representation
26What is the role of the decoder in diffusion-based models?
To reconstruct the image from its latent space representation
27What does latent space represent in image generation?
A lower-dimensional space capturing the abstract features of an image
28What is DALL·E?
An AI image generator released in 2021
29What is MidJourney?
An AI art generator released in 2022 specializing in artistic styles
30What is Stable Diffusion?
A 2022 open-source AI model for text-to-image generation
31What is Imagen?
A text-to-image diffusion model developed by Google
32What is Nano Banana (2025)?
A new AI image generator introduced in 2025
33What is LMARENA?
An AI image generation system mentioned as part of modern applications
34What makes diffusion models powerful compared to GANs?
They avoid issues like mode collapse and generate more stable, diverse outputs
35What role do transformers play in modern diffusion models?
They help model long-range dependencies and improve generation quality
36Why is cosine similarity used in CLIP?
To measure closeness between image and text embeddings in a shared space
37What real-world applications came from CLIP?
Image classification, image search, and multimodal reasoning
38What is a common challenge in GAN training?
Balancing the generator and discriminator to avoid collapse or overfitting
39What advantage do diffusion models have in interpretability?
They generate images step by step, making the process more understandable
40What innovation allowed text-to-image systems like DALL·E and MidJourney?
The combination of diffusion models and transformer-based architectures
41What does 'zero-shot' transfer, a capability enabled by CLIP, mean?
For example, a model trained on (image, text) pairs can 'classify' a 'dog' by seeing which prompt ('a photo of a dog' vs. 'a photo of a cat') it matches best, even if it wasn't trained on 'dog' as a specific label.