How Easy Is It to Remake a Video with AI? Reference Video + One Prompt

A child in a pumpkin costume dances in the frame, with a small figure wearing a black witch hat standing beside them. The wall has ghost-face decorations and bat ornaments, and the jack-o'-lantern is glowing.

This video looks like a live-action Halloween scene. But tweet author Mani says it was remade by AI.

Mani made it with Renoise Recast, running on Claude, with Seedance 2.5, a video generation model, as the underlying model. There were only two inputs: the original reference video and one prompt.

The prompt roughly meant: Mom says it’s Halloween time, but the kids are excited about something else—candy time.

Notice how the model understood this sentence. To kids, Halloween does not equal scary; it equals candy. In the output, there is no creepy haunted house—only a child in a pumpkin costume dancing on a wooden floor. The model grasped this layer of meaning, which is a more critical capability in video stylization than costume swapping.

Previously, achieving this kind of effect required frame-by-frame processing and maintaining character consistency, making the workflow long and the barrier high. Recast compresses the process into two steps: give a reference video and write a prompt. For ordinary users, the focus of AI video creation is no longer how to operate, but how to describe.

The comments are also discussing the result. One user called it “amazing,” another said “candy time is the real magic of Halloween,” and another, seeing the pumpkin-costumed character dancing in the frame, thought these little monsters were actually quite cute.