Can one CartVisual product request use two saved AI model references?

Separate a two-person cast from a two-person request

Consider an illustrative backpack campaign with two fictional adult characters, Rowan and Tessa. The team wants both to appear reliably across the campaign. A set of separate portraits and product scenes can serve that casting plan. A single picture of them exchanging the backpack introduces another requirement: two controlled identities and a coherent joint interaction within the same generated composition.

Write which result is actually needed before selecting inputs. If the campaign simply needs both characters represented, a coordinated pair may be sufficient. If the story depends on their interaction, two independent portraits cannot establish the hands, ownership and contact relationships of that joint scene. Do not describe the simpler output as though it records an exchange that never appears.

Use campaign casting to decide who belongs across the image set. That broader cast decision does not change the number of selected identity inputs in one request.

Check the selector rather than counting uploaded photos

In Product images, choose a purpose that actually uses the person reference. The current submission omits it for White and Details; selecting a saved model elsewhere does not make it apply to those outputs. For supported person scenes, the picker resolves one selected official or saved reference. Virtual try-on also uses a single selected reference in this workflow.

Selecting Tessa after Rowan replaces the selection for that workbench; it does not queue both identities for one image. The current image request carries one modelId when applicable. Product or garment uploads have their own evidence role. Adding Tessa’s photograph there does not create a second designated person slot or a rule assigning each face to its source.

Google’s Imagen customization documentation describes reference identifiers in its own API. In its person example, two reference images of the same person share one identifier, while a control image has another. That illustrates why image count and subject roles are different. It neither adds controls to CartVisual nor establishes a verified two-person production method for this guide.

Choose the output that matches the story

For the fictional launch, give R1 and T1 the same actual backpack edition and compatible creative direction. Review each person and product independently before assembling the two-panel graphic. Name which approved file supplies each panel. A correctly generated Rowan does not approve Tessa, and matching backgrounds do not prove a shared moment.

If the story requires a handover, the joint action has its own hand, strap and ownership relationships. Separate portraits or a panel arrangement do not establish that action. Do not describe the graphic as showing an exchange it never depicts. A second plausible figure that happens to appear in one output is also not proof that a second saved identity was followed.

Pilot each supported identity with the actual product

For R1, choose Rowan’s authorized reference in a supported product scene and review the current selection before submitting. Repeat separately for T1 with Tessa. Inspect each identity against its own source, then check the backpack’s label, seam pattern, straps and visible count. Keep the identity contact-sheet review separate from the decision about the two-panel story.

In the external handoff, record R1 → Rowan, T1 → Tessa, the shared product edition, and “coordinated two-panel campaign, no joint exchange depicted.” These identifiers are organizational names for fictional adults, not biographies, testimonials or reports of a real event. No portrait, product scene, API call, identity test or composite was created for this exercise.

Handoff the identity decision with the approved files

Give the external editor the separate approved images, character names used for organization, product edition and intended panel relationship. State whether the layout is a coordinated campaign comparison or a depiction of interaction. Avoid adding captions that invent a real biography, customer experience or testimonial for either fictional character. Their visual consistency and public role are separate review questions.

If a generated image happens to contain two plausible figures, do not treat that as evidence of two saved identities being controlled. Compare each figure to its authorized source and reconsider whether the unsupported second person belongs in the final asset. This example documents the current input boundary and production choice; it does not promise perfect identity retention, multiple-identity controls or a tested joint-scene result.