Contents
Reference-based generation works best when every reference has a job. Without roles, references can fight each other.
A reference can define identity, style, motion, composition, or the first frame of a scene. Treat each reference as a constraint for the model.
Use fewer, clearer references at first. Add more only when you understand what each one controls.
What you need
- One or more images, frames, videos, or library assets that define the desired look.
Step by step
- 1
Decide what must stay stable
Choose identity, product shape, style, camera, or motion as the main constraint.
- 2
Assign a role to each reference
Do not upload references without knowing what they should control.
- 3
Test one scene
Reach stability on a short scene before scaling into a full video.
- 4
Save the working setup
Add successful references and results to the library.
- 5
Reuse the setup
Bring the same assets into later scenes to keep the series consistent.
Reference types
Image reference
Useful for subject, product, character, pose, style, or lighting.
Start frame
Useful when the first visual state must be exact.
End frame
Useful when the scene needs a target state or transition.
Motion reference
Useful when movement matters more than exact appearance.
FAQ
Can too many references hurt the result?
Yes. More references can confuse the model if they do not share a clear role.
How do I keep a character consistent?
Use a strong identity reference, test short scenes, and save the best working frames for reuse.
