Product-in-hand AI video: references and review checks
Prepare the right reference images and a simple action so hands, packaging and scale remain reviewable in an AI product video.
By Janice Phil · Growth · Published
In this article
The hard part is the interaction
A product-in-hand AI video asks the generator to keep a face, hand and product coherent while they move together. The product can look correct in the opening frame and still bend, change label or slide through a finger later. Review the interaction over time rather than approving a single attractive still.
Use this format when holding the product helps explain size, grip or use. If the important information is on a tiny label, a separate stable product close-up may be clearer and easier to verify.
Give each reference a job
Provide a sharp image of the exact product and an authorized person reference. Where possible, include a reference for the desired grip or product placement. Keep the important label visible and describe which features must remain unchanged.
The package image below is already used in the roasOS demo. It illustrates the need for a clear product view; the copy on a package is not independently substantiated by its presence in a demo.
Choose one simple action
Start with a stable hold and a modest movement. A continuous sequence that opens a lid, removes an item, rotates the package and points at text introduces several chances for geometry to fail. Plan separate shots if those actions are all necessary.
Illustrative prompt: “The presenter holds the referenced package upright at chest height. Keep the front visible, fingers away from the label and the product proportions unchanged. Small natural body movement; no opening, rotating or changing the package. Use the approved spoken line.” Adapt this to the actual model controls rather than assuming it will enforce every constraint.
Review the contact points
Check the first, middle and last frames, then watch the movement at normal speed. Slow inspection catches intersections, while normal playback reveals unnatural motion and timing. Both views matter.
| Check | Accept when | Reject when |
|---|---|---|
| Grip | Fingers remain plausible and attached | Fingers merge or pass through packaging |
| Scale | Product size stays consistent relative to the hand | Container grows or shrinks |
| Label | Shape and key identity remain recognizable | Logo, text or colors transform |
| Motion | Hand and product move together | Product floats or slips independently |
| Speech | Approved line is understandable | Words or claims change |
Use a cutaway when the action is not reliable
If a model repeatedly fails on a necessary interaction, change the production plan. Use real product footage or a stable reference shot for the mechanism, then return to the presenter. This preserves the explanation without pretending a flawed generated action is accurate.
In roasOS, inspect the available Avatar or Custom mode inputs and selected model capabilities before generating. Keep the approved reference and rejection notes with the output so a later batch does not repeat the same failure.
A production pass does not approve the underlying claim. Confirm product facts separately and avoid presenting an AI actor as a real customer describing personal results.