audio alignment is off in later videos when calling pipeline.walk

[nateraw]

when you use multiple prompts/seeds (4 in my case), the audio for the later clips begin to be off from where it should be. This is actually because when calculating T, the audio offset calculation is incorrect:

https://github.com/nateraw/stable-diffusion-videos/blob/90039cf2448691d5ce3185ac33ddd0886606f713/stable_diffusion_videos/stable_diffusion_pipeline.py#L739

This should instead be running sum of num_interpolation_steps.

[nateraw]

Same issue on this line:

https://github.com/nateraw/stable-diffusion-videos/blob/90039cf2448691d5ce3185ac33ddd0886606f713/stable_diffusion_videos/stable_diffusion_pipeline.py#L753

[nateraw]

This logic should work as fix

num_interpolation_steps = [120, 150, 60]
fps = 30
audio_start_sec = 2.0
for i, num_step in enumerate(num_interpolation_steps):
     audio_offset = audio_start_sec + sum(num_interpolation_steps[:i]) / fps
     duration = num_step / fps
     print(f"Audio Offset: {audio_offset} | duration: {duration}")