Build a sketchy version of my video

Vikit-ai / sdk

Vikit.ai SDK let you develop easily video generators leveraging generative AI and other AI models.

https://vikit.ai

Apache License 2.0

49 stars 7 forks source link

Build a sketchy version of my video #6

Open jeffmac-aix opened 4 months ago

jeffmac-aix commented 4 months ago

Generating videos takes time, minutes to tens of minutes, even with parallelism and async patterns. I would love to see even a sketch y low fidelity version of my video, without music, without sound, but fast. Say my goal would be to get something to watch in less than 30 seconds. The latest part of the video could be generated a bit later as long as I am provided with a simple way to watch everything, even when not all videos chunks composing my final video are not generated yet.

jeffmac-aix commented 4 months ago

Using a new video building strategy would be the way to go:

no interpolation
explore how far we could play with image resolution / quality, or if other models would have been trained with lower resolution
No sound
Generating the first videos first: priority queue for video building

jeffmac-aix commented 3 months ago

try a variant not using ffmpeg and just generate the videos from platforms, and run them sequentially (may be even dynamically) from a video player capable of reading playlists, like VLC or the .js one

jeffmac-aix commented 2 months ago

https://replicate.com/chenxwh/t2v-turbo might be the type of model we could use for this, 6s generation for 2s video with low framerate