search

RedAIGC / Target-Driven-Distillation

Consistency Distillation with Target Timestep Selection and Decoupled Guidance
https://redaigc.github.io/TDD/
38 stars 4 forks source link

About training iterations #4

Open lovejing0306 opened 1 day ago

lovejing0306 commented 1 day ago

The training iterations in your paper are 20000, but in your codes, the training iterations are 100000. Can you tell me why? Xnip2024-09-29_16-48-56

palmerkuo commented 1 day ago

We tested some models from the later training stages, and they performed worse on certain quantitative metrics compared to the current model. The final released model corresponds to the one at 20k steps. @lovejing0306

lovejing0306 commented 8 hours ago

We tested some models from the later training stages, and they performed worse on certain quantitative metrics compared to the current model. The final released model corresponds to the one at 20k steps. @lovejing0306

ok, thanks. I have another question. Why do you set the cfg=3.5 in your paper?