Higgsfield Genjutsu vs Seedance 2.0 Mini vs MiniMax H3: can a kitten copy a man’s dance?

Motion transfer models take a reference video and put someone else into it. We wanted a small kitten, a 3-month-old about 40 to 55 cm tall on its hind legs, to do the viral car alarm dance that TV meteorologist Nick Kosir posted on TikTok in January 2024: a grown man in a grey suit hopping, kicking and swinging his arms next to a white SUV. On 2026-11-08 we gave the same dance and the same kitten photo to three models. The short answer: none of them did both things at once. Higgsfield Genjutsu (standard motion transfer, 480p) copied the dance almost move for move, but drew the kitten as tall as the man, with its helmet above the SUV’s tail lights. Writing the size into the prompt (40 to 55 cm, shorter than the tyre) and giving it a first frame with the kitten at the right size changed nothing. Genjutsu’s API has no size setting, and the open-source copy Openjutsu works by turning the original dancer into a mannequin of the same outline, which would explain why the new character fills a human-sized shape. Seedance 2.0 Mini (reference-to-video, 720p, the full 12 seconds) kept the kitten small in every frame when we passed the small first frame as a reference image. The man never appeared and the camera and timing followed the original, but the kitten swapped his kicks and arm swings for generic bouncing and ran on all fours where he crouched. Describing every move second by second in the prompt made it stiffer, not closer. MiniMax H3 (multimodal reference, 768P, 5 seconds; it bills the reference video’s seconds as well as the output’s) also kept the kitten small, and also lost the moves: where the man hops on one leg with an arm over his head, the kitten walks toward the camera with its paws out. Our reading: these models read the moves from the reference dancer’s body. When the new character is shaped like the old one, the moves carry over; when a kitten has to copy a grown man, a model either keeps his body size or keeps only his rhythm. Getting the first frame right was easy by comparison: an image model needed three tries, and the size only came out right once we anchored it to the car (the top of the helmet at the wheel hub), not in centimetres. Through the APIs, each test clip cost between $0.58 and $0.71.