You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
您好,我用rife_m,修改了ifnet_m中如上您给的代码,然后参照论文Fig8做的图片I0 I1(基于demo里的两张图片进行了左右侧裁剪),出来的中间帧图片是一片混乱的像素,可以帮忙看下什么问题吗?
rife_m推理如下:
img_list = [img0]
p = 2 ** args.exp
for i in range(p-1):
t = (i+1)*(1./p)
print(t)
img_list.append(model.inference(img0, img1, timestep=t))
img_list.append(img1)
Activity
hzwer commentedon Nov 23, 2022
在 model/IFNet_m.py 里面把 timestep 改写一下
这样应该就创建了一个渐变的 temporal encoding
qijiaojiao commentedon Nov 28, 2022
您好,我用rife_m,修改了ifnet_m中如上您给的代码,然后参照论文Fig8做的图片I0 I1(基于demo里的两张图片进行了左右侧裁剪),出来的中间帧图片是一片混乱的像素,可以帮忙看下什么问题吗?
rife_m推理如下:
img_list = [img0]
p = 2 ** args.exp
for i in range(p-1):
t = (i+1)*(1./p)
print(t)
img_list.append(model.inference(img0, img1, timestep=t))
img_list.append(img1)
hzwer commentedon Nov 28, 2022
我试了一下,用目前的 RIFEm 模型,对这两张图片做插帧会出一片蓝色,具体原因可能是模型无法适应镜头移动,我猜测训练的时候如果加一些镜头移动的增广能修复,很抱歉我之前没有注意到
改用 practical RIFE 里的模型,用 v4.6 模型 https://github.com/hzwer/Practical-RIFE#usage ,可以得到正常的结果
输入(左右各裁剪 32 pixel):


输出:

我的 timestep 生成方式:
qijiaojiao commentedon Dec 1, 2022
谢谢了!