Higgsfield have multiple models available, people use Kling usually 2.5 & 3. There are a few good examples posted right now you'll notice the subtle differences.
I have tried to generate things myself and it's extremely hard to have more than 7-8 clips that are consistent, eventually you'll accept a compromise. I think it's why there isn't any long form content being done yet. Getting good results is sometimes just "chance" regardless of how many reference data you have.
I've been doing some content with people at https://industrialallusions.com