INNER CODE UNIT · Python
times
jdh-algo/JoyHallo · scripts/app.py:345
times = audio_emb.shape[0] // clip_length
tensor_result = []
# generator = torch.manual_seed(42)
generator = torch.cuda.manual_seed_all(42) # use cuda random seed
for t in range(times):
print(f"[{t+1}/{times}]")
if len(tensor_result) == 0:
# The first iteration
motion_zeros = source_image_pixels.repeat(
cfg.data.n_motion_frames, 1, 1, 1)
motion_zeros = motion_zeros.to(
dtype=source_image_pixels.dtype, device=source_image_pixels.device)
pixel_values_ref_img = torch.cat(
[source_image_pixels, motion_zeros], dim=0) # concat the ref image and the first motion frames
else:
motion_frames = tensor_result[-1][0]
motion_frames = motion_frames.permute(1, 0, 2, 3)