A coffee cup is easy. That’s what everyone assumes. Right up until shot three, where the cup has grown a second handle, the logo has melted into abstract art, and the lid has become a straw. Props look like the simplest part of AI video — no face, no performance, no dialogue — and then they quietly wreck more product shots, brand deals, and story beats than any other single element. A mug that changes shape between cuts reads as a mistake to any viewer. Even one who couldn’t tell you why the scene felt cheap.
Objects obey their own consistency rules, and they’re different from the rules for people. Faces have reference features built into the tools. Your client’s sneaker does not. This guide covers how I keep products, cups, cars, and story props stable across a sequence: the reference shot that actually works, how to name materials and logos so the model gets them right more often, where props physically melt first — and the honest cases where the smart move is to composite the real product in the edit and stop fighting the generator.
Objects Are Characters With Worse Agents
It helps to think of a recurring prop as a cast member. One with no agent. No reference slot in most tools. And no fan base demanding it look right. People get consistency features because people are what the tools are sold on. Objects get whatever the model’s general training gives them, which is: a plausible version of the category. Ask for “a red sports car” ten times and you get ten red sports cars — each a slightly different interpretation. That’s fine for a background vehicle in one shot. It’s a disaster for a car the story keeps returning to.
Three traits make objects trickier than they look:
- Geometry is unforgiving. A face can drift a few percent and still read as the same person. A product that drifts a few percent in its proportions reads as a different model of that product. Viewers know what a specific phone, bottle or car looks like, and hard edges give drift nowhere to hide.
- Text and logos are the first casualty. Current models still mangle small text. A logo that survives a wide shot will smear in a close-up, and a label with real words on it will come back as convincing-looking nonsense. If the brand matters, plan for that from the start, not in the edit bay.
- Scale relationships slide. The cup that was correctly sized against a hand in shot one is a soup bowl in shot four. Models are loose with relative size unless something anchors it — usually a hand, a table edge, or a person holding the object in the reference frame itself.
The mindset shift is simple: anything that appears in more than two shots gets the character treatment. It gets a reference, a written spec, and a review pass. The same system thinking behind character consistency in AI video applies — references, fixed wording, checking — just retuned for things that don’t blink.
The Prop Reference Shot You Need
Most prop references fail because they’re photographs of the object in a scene: on a desk, in someone’s hand, half in shadow. That is a mood photo, not a reference. A working prop reference is boring on purpose. Closer to a catalogue shot than a lifestyle image.
What actually works, per prop:
- Neutral background, full object visible, nothing cropped. The whole silhouette needs to be in frame with margin around it. Cropped edges teach the model to guess what the rest looks like, and it will guess differently each time.
- Flat, even light. You want the true colour and material readable. Dramatic light in the reference bakes shadows into the object’s “identity,” and the model then reproduces those shadows as if they were paint.
- Two or three angles for anything with real geometry. A cup needs front and three-quarter. A car needs front three-quarter and side profile at minimum, ideally a rear three-quarter too. A phone needs front, back, and the edge profile if its thickness matters to the shot.
- One “in hand” or in-context frame for scale. A single reference of the object being held, placed, or worn at correct scale anchors proportions in a way pure product shots do not.
So where do these come from? For fictional props, generate them deliberately as stills first — one clean generation session just for the object, reviewed like you’d review a character sheet, before it ever appears in a video prompt. For real products, photograph the actual item on a plain background if you can. A real photo reference beats a generated one whenever the product has to be the product.
On Seedance 2.5, feed the prop references in alongside the character references and name their job in the prompt: “use reference 3 for the backpack.” On Veo, the reliable route for prop-critical shots is image-to-video — build the still frame with the prop correct, then animate. The deeper craft of choosing and feeding references is covered in the guide to reference images for consistent characters, and every word of it applies to objects too.
Naming Materials, Colours And Logos In Prompts
When you describe a prop in text, the model needs the same things a prop buyer would need: exact colour, material, finish, and the details that make this one different from the generic version. “A water bottle” buys you any water bottle. “A matte forest-green steel water bottle, black screw cap, no label” buys you the same one repeatedly. See the difference?
The descriptors that earn their place:
- Material plus finish. Brushed steel, matte plastic, glossy ceramic, worn leather, frosted glass. Finish changes how light hits the object, which is most of how viewers recognise it across shots.
- Colour with a qualifier. Bare colour words drift. “Forest green,” “off-white,” “gunmetal grey,” “faded denim blue” all narrow the range. Pick the specific term once and paste it forever, unchanged.
- Countable features. Two handles or one. Three buttons. A single stripe down the side. Counts are things you can also check in review, which makes them doubly useful.
- The distinctive marker. Every recurring prop should have one feature you could recognise in a thumbnail: the red cap, the cracked corner, the unusual shape. Distinctive props drift too, but you spot it instantly instead of three shots later.
Logos deserve their own paragraph, because they cause outsized grief. If the object carries a logo or label text, assume generated versions will be wrong — subtly at distance, badly up close. Your options, in order of sanity: keep logo-bearing shots wide or in motion, so the mark reads as a shape rather than text; generate the shot with a clean, unbranded version of the object and add the real logo in the edit as a tracked overlay in Premiere Pro; or composite the real product entirely, which is the next section’s topic. What does not work? Generating twenty takes hoping one spells the brand name correctly. That lottery is not a workflow.
Whatever wording you settle on, lock it. The prop gets a spec string just like a character does — the identical-words discipline from series-level character consistency is what keeps the “same” object the same object in episode six.
Contact Points: Where Props Melt First
Watch generated footage closely and you learn something: props rarely fail in isolation. They fail where they touch something. The rim of the cup where lips meet it. The handle where fingers wrap it. The phone where a thumb should be pressing. Contact points are where the model’s understanding of two objects at once gets stress-tested — and hands, already the hardest thing in AI video, make it worse.
The recurring failure modes:
- Grip drift. Fingers pass through the handle, hold the cup by its rim, or wrap an object that then changes size to fit the grip. The hand and the prop are negotiating in real time and neither wins cleanly.
- Weight loss. Objects that should press, dent, hang or swing float instead. A heavy bag carried by a weightless strap, a book that does not bend the hand holding it. Nothing looks digital faster than a prop with no physics.
- Surface merge. A glass on a table blends into the tabletop at the base. A car tyre melts into wet tarmac. Where object meets surface, edges go soft and shapes fuse.
- Reflection chaos. Glossy and transparent props — bottles, glasses, car paint, phone screens — reflect an environment that does not match the scene, or change their reflections every few frames.
Practical counters that actually help: keep critical contact brief. A hand picking up the product, holding it steady for the beat, and setting it down generates far more cleanly than a long take of fingers manipulating it. Frame the money moment — the product held up, label to camera — as a held pose rather than mid-action. Generate the pickup and the put-down as separate short clips and cut between them; the edit implies the handling you never had to render. And when a hand must do something complex with a prop, slow the action down in the prompt. Slow movement gives the model fewer chances to glitch per second of screen time — a principle that shows up across temporal consistency work generally.
When To Composite The Real Product In The Edit
Here’s the part nobody selling AI tools likes to say: for real commercial products, the most professional move is often to stop generating the product at all. Generate the scene — the person, the setting, the light, the motion — and composite the actual product photograph into the shot during the edit. This is not a failure of the AI workflow. It’s how a lot of polished commercial work quietly gets made.
Composite the real product when:
- The product is the ad. A client’s actual bottle, shoe or device, where the label must be exact and the shape is legally and commercially non-negotiable.
- The shot is a close-up. The closer the camera, the worse generated geometry and text get, and the easier a clean product cutout is to place convincingly.
- The product is mostly static in frame. Sitting on a table, held steady toward camera — these composite beautifully with a tracked still, a matched shadow, and a touch of grain. A product tumbling through the air does not.
The craft is in the blend. In Premiere Pro, the composite lives or dies on four things: matching the scene’s light direction on your product photo (shoot or choose the product image lit from the same side), adding a soft contact shadow where it sits or is held, matching grain and sharpness so the product isn’t suspiciously cleaner than the generated footage around it, and a colour pass so its white balance belongs to the scene. Skip the shadow and the product floats. Skip the grain match and it looks pasted. Do all four and most viewers will never question it.
The hybrid is usually the winner: generated versions for wide and moving shots, where drift hides in distance and motion; the real product composited for the hero close-ups, where it matters. The audience gets a product that looks perfect exactly when they’re looking at it hardest. Which is, of course, the only time it matters.
None of this is complicated. But it has to be done on purpose. Give recurring props references, written specs, and review time — exactly like cast members. Respect the contact points. And when the object on screen has to be the real thing, be honest about it, and put the real thing there in the edit.