SDXL and Cascaded Diffusion Quiz

Answers and explanations appear as you take the quiz.

1. What native resolution does SDXL generate images at, compared to the original Stable Diffusion?
   • 1024x1024 vs 512x512
   • 256x256 vs 128x128
   • 2048x2048 vs 1024x1024
   • 768x768 vs 256x256
2. What is the role of SDXL's refiner model?
   • It encodes the text prompt
   • It adds fine detail during the final low-noise denoising steps
   • It compresses images into latent space
   • It generates the initial random noise
3. How many text encoders does SDXL use?
   • One
   • Three
   • Two
   • Zero
4. What is the core idea of cascaded diffusion?
   • Training a single giant model on all resolutions at once
   • Removing the text encoder to speed up generation
   • Replacing diffusion with a GAN at every stage
   • Chaining a base model with super-resolution diffusion models that progressively upscale
5. Why does SDXL denoise in a latent space rather than directly on pixels?
   • It makes denoising computationally cheaper
   • Latent space images look more colorful
   • Pixels cannot be denoised
   • It avoids needing a text prompt
6. Which technique reduces SDXL to as few as one to four generation steps?
   • Crop conditioning
   • Distillation methods like SDXL Turbo and Latent Consistency Models
   • Adding more text encoders
   • Increasing the VAE size
7. What do SDXL's size and crop conditioning inputs tell the model?
   • Which training dataset owns the prompt
   • The number of denoising steps to run
   • The target image size and framing or crop
   • The image's final file format and compression level
8. What problem does noise-conditioning augmentation address in a cascaded diffusion pipeline?
   • It removes the need for a text prompt
   • It guarantees that every stage produces the same resolution
   • It converts the latent image into an audio signal
   • It helps later super-resolution stages handle errors or uncertainty in the previous stage's low-resolution image