search

cswry / OSEDiff

80 stars 0 forks source link

Question: Is the time step directly fixed to T in the input of diffusion (the largest time step) #5

Open Feynman1999 opened 2 weeks ago

Feynman1999 commented 2 weeks ago

Is the time step directly fixed to T in the input of diffusion (the largest time step)

cswry commented 2 weeks ago

yes

Feynman1999 commented 2 weeks ago

Have you ever tried ablation experiments on the setting of t. Intuitively, an LR contains a lot more information than xt when T=1000. Is it because of the presence of Lora weight that T becomes less critical?

Feynman1999 commented 1 week ago

Can you share your understanding of time step settings， thanks!!

cswry commented 1 week ago

Hi, A larger T tends to generate more, while a smaller T tends to preserve fidelity.

T balances fidelity and perception. We aim for the better generation performance of OSEDiff, so we chose 999.

cswry commented 1 week ago

Have you ever tried ablation experiments on the setting of t. Intuitively, an LR contains a lot more information than xt when T=1000. Is it because of the presence of Lora weight that T becomes less critical?

Have you ever tried ablation experiments on the setting of t. Intuitively, an LR contains a lot more information than xt when T=1000. Is it because of the presence of Lora weight that T becomes less critical?

Your judgment is correct. Setting different T values at the input of UNet only affects the convergence speed, with little impact on the final model performance.