how to extend real video?

[loboere]

Any nodes or workflows to extend a real video instead of a generated one?

[DragonDiffusionbyBoyo]

just take the last frame off the real video and use that as the first one for a generation. Simple. Keep using first and last frames to extend your videos into minutes. This is how it is done.

[loboere]

But a frame does not obtain information from what the previous video contains, the idea is that it takes the information from the previous video to extend all the physics, movement, lighting of the previous video.

[DragonDiffusionbyBoyo]

just use a colour match node in comfyui. I have been doing this for years. with the correct prompts you can take the last frame and extend it. This is all the video extender is doing underneath the hood but doing the colour matching for you under one prompt. Extending in sections of 6 seconds allows you to vary the prompts and get different actions rather than hoping for a hollywood blockbuster off god knows how long from one prompt. I will let you decipher what I have said and then you will understand.

[loboere]

The idea is for the extender to know what's happening in the previous video, not from the last frame, but from all the frames in the previous video.

[DragonDiffusionbyBoyo]

you would need to use a vae encoder to turn the video into latents. i find the extender too slow and my method you can have 3 or 4 outputs in the time taken and chose the best. You would also need to adjust the frame rate to the native rate of the model as well as correctly sizing it. If you are looking to shove a 30 second video in and add another 10 or seconds you can forget it as you will not have enough Vram