https://www.reddit.com/r/StableDiffusion/comments/11yejrj/another_temporal_consistency_experiment_the_real/


무슨 원리인지는 잘모르겠는데, 쩌네


(수정)

방법이 댓글로 적혀있어서 일단 메모용으로 써둠

This is the text of it...

  1. Take your video clip and export all the frames in a 512x512 square format. Here I chose my doggy and it is only 3 or 4 seconds.

  2. Look at all the frames and pick the best 4 keyframes. Keyframes should be the first and last frames and a couple of frames where the action starts to change (head turn etc, , mouth open etc).

  3. Copy those keyframes into another folder and put them into a grid. I use https://www.codeandweb.com/free-sprite-sheet-packer . Make sure there are no gaps.

  4. Copy the grid photo into ControlNet (see screengrab) and ask Stable Diffusion to do whatever. I asked for a Zombie Dog.

  5. When you get a good enough set made, cut up the new grid into 4 photos and paste each over the original frames. I use photoshop. Make sure the filenames of the originals stay the same.

  6. Use EBsynth to take your keyframes and stretch them over the whole video. EBsynth is free. My settings are in a screengrab.

  7. Run All. This pukes out a bunch of folders with lots of frames in it. You can take each set of frames and blend them back into clips but the easiest way, if you can, is to click the Export to AE button at the top. It does everything for you!

  8. You now have a weird video.