ComfyUI_video-image-motion-transfer

Introduces a node that tries to approximate the entire video using it's first frame (that we stylize) by warping it using optical flow extracted from the video. First we do image-to-image using the reference video's first frame and a depth controlnet. The generated object doesn't have to closely resemble the reference like on the demo. Then the generated image and the video frames are fed into the node and it returns a warped video. No ai models are used by the node. The example workflow uses SDXL but you can get it to work with any arch if you manage i2i.