Diffusion Transformers Quiz

Answers and explanations appear as you take the quiz.

1. What component does a Diffusion Transformer replace compared to traditional diffusion models?
   • The convolutional U-Net denoiser
   • The text encoder
   • The image decoder
   • The noise schedule
2. How does a DiT turn an image into something a Transformer can process?
   • It converts pixels directly to words
   • It splits the latent image into patches treated as tokens
   • It uses one token for the whole image
   • It converts the image to audio first
3. What did the original DiT paper find about scaling?
   • Performance plateaus quickly with size
   • Bigger models always perform worse
   • Performance improves predictably as model size grows and patch size shrinks
   • Patch size has no effect
4. How do DiTs typically inject conditioning like the timestep and text prompt?
   • By appending it as extra pixels
   • By retraining the model each step
   • Via a separate convolutional branch
   • Through adaptive layer normalization (adaLN)
5. What does the 'adaLN-zero' initialization accomplish?
   • It makes each Transformer block start as an identity function for stable training
   • It deletes the normalization layers
   • It freezes the entire model
   • It zeroes out the input image
6. Which well-known system uses a Transformer backbone over spacetime patches for video generation?
   • A traditional GAN
   • OpenAI's Sora
   • A recurrent neural network
   • A decision tree model
7. Why can full self-attention become costly when a DiT processes many patches?
   • Its cost stays constant for any token count
   • It must retrain the U-Net at each denoising step
   • Its pairwise token comparisons grow quadratically with the number of tokens
   • It requires a separate language model for every patch
8. After a DiT processes latent image patches with its Transformer blocks, what happens to the patch outputs?
   • As one text token for an entire video
   • As only a sequence of audio waveforms
   • As a separate convolutional U-Net for each frame
   • They are reassembled and decoded back into pixels