zsyOAOA / ResShift

ResShift: Efficient Diffusion Model for Image Super-resolution by Residual Shifting (NeurIPS 2023 Spotlight)
Other
823 stars 42 forks source link

steps设置为15 #62

Open UserYuXuan opened 4 months ago

UserYuXuan commented 4 months ago

请问在论文中inpainting中steps设置为4,但论文中steps还可以设置为15,设置为15时是否修复的效果会比4时更好呢?

zsyOAOA commented 4 months ago

在我们的会议论文(只包括SR)中, steps设置为了15.

在扩展的期刊论文中,我们尝试了SR, inpating, Face restoration三个任务,steps都进一步加速到了4. 在inpainting上,steps=4应该是最优的,因为训练的时候steps也固定到了4,将steps调整到15的话,我并没有尝试过,不确定效果如何。

UserYuXuan commented 4 months ago

在我们的会议论文(只包括SR)中, steps设置为了15.

在扩展的期刊论文中,我们尝试了SR, inpating, Face restoration三个任务,steps都进一步加速到了4. 在inpainting上,steps=4应该是最优的,因为训练的时候steps也固定到了4,将steps调整到15的话,我并没有尝试过,不确定效果如何。

十分感谢

UserYuXuan commented 4 months ago

在我们的会议论文(只包括SR)中, steps设置为了15.

在扩展的期刊论文中,我们尝试了SR, inpating, Face restoration三个任务,steps都进一步加速到了4.在inpainting上,steps=4应该是最优的,因为训练的时候steps也固定到了4,将steps调整到15的话,我并没有尝试过,不确定效果如何。

UserYuXuan commented 4 months ago

Snipaste_2024-05-02_23-55-39 请问在训练自己的数据集,inpainting任务时出现第4步的loss总是比第三步高,这是正常现象吗?

zsyOAOA commented 4 months ago

Snipaste_2024-05-02_23-55-39 请问在训练自己的数据集,inpainting任务时出现第4步的loss总是比第三步高,这是正常现象吗?

是的 @UserYuXuan