Some questions about section 3.4 of the paper
#45
by johnn95 - opened
1.Are D′(xt−ns, t − ns, c) and D the same model ?
2.Is the input condition c empty?
3.Is the training goal to make the student's unconditional output close to the teacher's unconditional output?
4.What are the numbers of image for training the conditional objective and finetune with unconditional objective?
Looking forward to your reply！
- Not the same model
- c is never empty. Unconditional refers to the condition on x_t, not c.
- Yes.
- We used different iterations for different stages. Earlier stages converge faster. About 10k iterations more or less for each stage.
Thank you for your reply
johnn95 changed discussion status to closed