I might upload a photograph because I want its jacket, not its lighting or its background. A second image might be there for colour. A voice note might describe pacing rather than supply dialogue.
A single upload box doesn’t explain those intentions. The tool needs some way to ask what each reference is for, and to preserve that answer through later steps.
I prefer named roles: character reference, style reference, composition sketch, narration. They make corrections specific. If the character looks right but the lighting is wrong, I should be able to change the style reference without replacing the character.
Outputs need similar separation. A video export is convenient for watching, but an editable project needs its shots, audio and timing kept apart. Changing a sentence should not automatically mean resampling every visual.
The models are improving quickly. The part I keep thinking about is the workspace around them: where the references live, which decisions are saved, and how much of the result survives the next revision.