Trying out Seedance 2.5 for text to video

I have heard a lot of hype about Bytedance's latest text to video model - Seedance 2.5. I bought some credits yesterday and gave it a go. The process of buying credits was not straightforward. When I searched for the website there seemed to be a bunch of noise - other websites with similar names. This goes along with the hype I guess, but when you are handing over a credit card this is a bit of a pain. I asked chat gpt for the url. This seems very backward, llms can easily get this stuff wrong, but I got to the right place in the end. The base credits package was about $30 which is perfectly fine for an initial trial. And those credits last 90 days.

Seedance takes a set of references. I just used photos, but you can use a wide variety of them from audio to clay renders from Blender. My son is a big fan of blender and has recently made a model of a steam train. He exported a nice render of this and we used it as the only reference. Initially I had wanted to add pictures of my 2 kids and put them in the video, but this gave an error because they are real people. Not sure what would constitute a sufficiently fake person to get past this filter, but for now I just removed them.

My prompt was "A boy playing with lego while a slightly older girl does Irish Dancing beside him". I did not mention the train, but it appeared in the video anyway which is just what we wanted. This is the video https://youtu.be/UFNJ7tPVfQw?si=TgyDEocmkspR2sGy. And my son did want laser eyes, so that is spot on. Although the subject is silly the quality of the video is excellent. It took about 3 minutes to generate this 4 second clip and it cost about 50 cent.

This model is not yet on the Artificial Analysis leaderboard for Text to Video, but will almost certainly be near the top. I have plenty more credit left to use, so will probably try to make some more videos soon.