MiniMax H3 reference-to-video binds a person from one or more pictures into a generated clip. It is very good at it, and it is a little too good at it: the model treats the whole reference as information, not just the face. The most common support question about the Flow and Portrait templates is some version of "why is my character on a yacht standing in her bathroom?" This lesson is about controlling what the model takes from the reference.
Academy / Video on one GPU / Lesson 2
Pro lesson · 6 min
Reference-to-video without background bleed
H3 copies more from the reference than the face. If the reference has a kitchen in it, your beach clip gets a kitchen. How the reference is read, how to crop it, and how to write the scene so it wins.
The rest of this lesson is for Pro members
Pro is $29 for 3 months and includes every lesson, every premium workflow and the whole character roster.