Guide-linked quiz · Medium Level
Classifier-Free Guidance Quiz
Review classifier-free guidance, conditional dropout, quality-diversity tradeoff, inference costs, and guidance scale. Check scope and assumptions.
Question 1 of 8
Why do diffusion models apply classifier-free guidance?
Preview every question8 questions
Answers and explanations appear as you take the quiz.
- Why do diffusion models apply classifier-free guidance?
- To strengthen how closely the output follows the conditioning prompt
- To compress the model into a smaller size
- To remove the text encoder entirely
- To convert images into audio
- How does a model learn to make unconditional predictions for classifier-free guidance?
- By training a second separate classifier
- By randomly dropping the conditioning during training
- By using only black images
- By disabling the GPU
- What does increasing the guidance scale generally do?
- Always improves diversity
- Removes the prompt's influence
- Strengthens prompt adherence but can cause oversaturation
- Makes the image grayscale
- Roughly how many forward passes per denoising step does standard classifier-free guidance require?
- Zero
- Ten
- One hundred
- Two (one conditional, one unconditional)
- What advantage does classifier-free guidance have over classifier guidance?
- It needs no separately trained classifier
- It requires more training data
- It only works on text models
- It eliminates the diffusion process
- In the common classifier-free guidance formula, what does guidance scale s=0 produce?
- The conditional prediction with extra amplification
- The unconditional model prediction without conditioning
- A prediction from a separately trained classifier
- A sample with all denoising removed
- How does classifier-free guidance combine model outputs at inference?
- It combines two image classifiers
- It removes the conditional prompt
- It combines conditional and unconditional score estimates
- It adds a learned reward model after decoding
- What does a guidance scale of 1 generally correspond to in the common formulation?
- The unconditional prediction only
- A classifier-based gradient
- A zero-noise image
- The conditional prediction without extra guidance amplification
Keep testing yourself
More quizzes picked for your level.