how to make inference on ted dataset

[Chromer163]

hello， thanks for excellent work! I wander how to make inference on Ted dataset, when I try default setting like Vox using Ted pretrained model, the generated video is terrible, may you show some examples on it? Thanks!

[Dheeraj791]

+1

[Suvi-dha]

Make these changes in your demo.ipynb code:

import torch
import os
os.environ['CUDA_VISIBLE_DEVICES']='3'
# edit the config
device = torch.device('cuda:0')
dataset_name = 'ted' # ['vox', 'taichi', 'ted', 'mgif']
source_image_path = '####'
driving_video_path = '#####'
output_video_path = './generated.mp4'
config_path = 'config/ted-384.yaml'
checkpoint_path = 'checkpoints/ted.pth.tar'
predict_mode = 'relative' # ['standard', 'relative', 'avd']
find_best_frame = False # when use the relative mode to animate a face, use 'find_best_frame=True' can get better quality result

pixel = 256 # for vox, taichi and mgif, the resolution is 256*256
if(dataset_name == 'ted'): # for ted, the resolution is 384*384
    pixel = 384

[ulucsahin]

Any updates to this? Even with config suggested above, results are very bad.