Dime Dime Baby AI Reference Clip Test: What Three Samples Taught Us
We made three five-second Dime Dime Baby AI samples from one photo and one dance reference each. Here is what the reference controlled, what drifted and how we fixed it.
What we tested
We made three samples, all with the Seedance 2 Mini model, at 480p, 9:16 and five seconds. Each used one photo for the person or pet and one short dance clip as a motion reference. You can watch them on their own pages: the bike-ride sample, the garden-party sample and the golden retriever sample. They are team samples, not customer results.
Each took about two minutes to render.
What the photo and the reference each control
- The photo supplies who appears: the face, hair, clothing and, for the dog, the fur and paws.
- The reference clip supplies what happens: the order of moves, the camera and, unless you ask for a change, the scene.
In the first sample the reference was the first five seconds of a wedding entrance dance: bite some food, give the camera a brief serious look, then turn away. The result kept that order with a new person in it. The moves follow the reference in sequence but not frame by frame.
Four things that drifted, and what fixed them
1. Clothing changed. In our first attempt the man’s plain white round-neck T-shirt came back as a collared top, and his expression was too cheerful. Naming the garment exactly (a plain round-neck T-shirt with no collar and no buttons) and describing the expression (a neutral stare after the bite) kept both in the second attempt.
2. Parts the photo does not show were invented. The portrait of the woman did not show her legs, so the model had to guess. We named dark jeans and plain sneakers in the prompt. If something is out of frame, say what it should be.
3. Changing the scene works, but the camera moves. For the second sample we asked for a garden party at sunset instead of the original hall. The new setting appeared, and the camera ended up slightly tighter than in the reference. Treat a scene change as a remix, not a controlled test.
4. A pet cannot copy a human gesture literally. The golden retriever sample used a clip of an aunt raising a hand to her mouth. We described it as an empty paw coming up near the muzzle, kept the dog’s face, fur, paws and tail, and told the model not to copy the aunt’s face, grey hair or dress.
How to judge a result
Watch the full clip, not a still. Check the clothing against your photo, the face across several moments, and whether the hands and mouth stay clear when the bite happens. Our stills looked right before the full playback showed small differences, and a likeness judgement needs the whole clip.
What this means for your own video
- Choose a photo with a clear subject and space for the body, as in our photo guide.
- Write clothing and expression in plain words, and describe anything the photo does not show.
- Ask for a new scene only if you accept that the framing may change.
- Use the structure in the dance template guide and the wording in the prompt guide.
When you are ready, make your own video.