Basics · 2/5
Lock your product with references
The difference between an AI ad for a cup and an AI ad for your cup is one node and one role.
What you can do after this
Keep your actual product and creator recognizable across every shot in a run.
Nodes used
- reference
- generate_image
- generate_video
Before you start
- Finish "Your first AI video", because this lesson edits that same graph.
In this lesson
- 0:00
Add the product to your library once
A product is a workspace entity with a packshot, not a file you re-upload per run. Add it once and every future run can reach for it.
- 0:10
Drop a reference node on the canvas
The reference node resolves a library entity into reference images. Picking your product fills its images output with the packshot, tagged with the role "product".
- 0:20
Understand what a role does
Roles tell the model what a reference is for. identity is a face that must repeat, product is an object that must stay identical, scene is a location, style is an outfit or look.
- 0:30
Wire references into the keyframe
The references port on generate image accepts several inputs and keeps their order, which is the order referenceRoles is read in.
- 0:40
Let the prompt describe the scene, not the product
Once the packshot is carrying the product, the prompt should stop describing it. Spend the words on light, surface, and camera instead.
- 1:15
Check that the lock survives motion
Wire the keyframe into generate video as before. Because the identity is settled in the still, the motion step has far less room to invent.
TranscriptThe same lines are burned into the video as subtitles.
- 0:00A prompt alone gives you a generic cup. This pins your real one.
- 0:05Open the library and add your product once — name, packshot, done.
- 0:10Back on the canvas, add a reference node.
- 0:15Pick the product. It resolves the packshot with role 'product'.
- 0:20Roles matter. product means: this must stay identical.
- 0:25identity, scene, style and product are separate roles.
- 0:30Wire the reference node's images output into generate image.
- 0:35The references port takes several inputs, in order.
- 0:40Now the prompt describes the scene, not the product.
- 0:45"On a cafe counter at golden hour, warm rim light, shallow depth."
- 0:50Run it. The label, the shape, the colour come from your packshot.
- 0:55Compare it against the prompt-only version from lesson one.
- 1:00Same direction, but now it's your product, not a lookalike.
- 1:05Add a creator reference the same way for a face that repeats.
- 1:10Each reference is reusable across every run in the workspace.
- 1:15Wire the keyframe into generate video as before.
- 1:20The lock holds through motion — that's keyframe-first working.
- 1:25One reference, set once, and every shot stays on brand.
- 1:30Next: what to do when one shot comes back wrong.
Frequently asked questions
How many references can one keyframe use?
The references port is a multi-input, so you can wire in several: typically a creator identity plus a product, and sometimes a scene plate. Their edge order is the order referenceRoles is applied in, so keep identity first when a face is involved.
Why does my product still drift after adding a reference?
Almost always because the prompt is still describing the product and fighting the packshot. Once a reference carries the object, delete the product adjectives from the prompt and spend the words on lighting and camera instead.
