I spent three generations getting the knight’s face right on Lost Garden. Same jaw, same scar placement, same tired eyes. Locked. Then I looked at the medallion around his neck across those same three shots and it had three different engravings on it.
Nobody warns you about this part. AI video models can now hold a face steady across a scene, but the object in a character’s hand is still treated as a sentence, not a subject, and sentences drift. That’s the answer, and it’s also the fix: give the prop the same treatment you already give the character. A locked reference, built once, reused everywhere.
Why does a locked character still have a prop that keeps changing?
Because the two problems aren’t solved by the same mechanism. Character consistency in 2026 mostly comes from identity-preserving reference features: Runway’s Gen-4 References, for instance, let you feed in up to three reference images and the model encodes them at what Runway’s own help documentation calls the entity level, so a face, a body type, or a specific object can be treated as a locked unit rather than re-described from scratch. The catch is that most people only ever point that feature at the character.
The sword, the medallion, the cup, the battered lantern someone carries through six scenes: those usually stay in the text prompt. And a text description is a suggestion, not a lock. “A silver medallion with an engraved symbol” means something slightly different to the model every time you type it, even if you type the exact same words, because the model isn’t retrieving an object, it’s reconstructing one from a category.
Character consistency and prop consistency are solved by different habits, and only one of them gets built by default
How do you actually build a prop reference sheet?
The same way you’d build a character sheet, just for the object instead of the person.
- Generate the prop alone, in image generation, from two or three angles, with nothing holding it. A held prop introduces its own inconsistency, because the model is now also guessing at hand position, grip, and partial occlusion at the same time it’s trying to lock the object’s shape.
- Lock the sheet before any video generation starts. If the prop looks wrong in the reference image, fix it there. Don’t try to correct it shot by shot once video generation has already begun; you’ll just get six more variations instead of one right answer.
- Attach that sheet as an actual reference asset, not a text description, using whichever mechanism your video tool gives you for it.
- Keep the descriptor language identical in every prompt that follows, even after the reference is attached. Models still lean on the words around the reference.
- Combine the prop with the character at shot time, not before. Character turnaround and prop turnaround stay separate documents that only meet inside the actual shot prompt.
Four steps to lock a prop before you generate a single shot with it
A prop that evolves across a story, a lantern that goes from lit to dark, a ring that gets damaged, needs its own updated reference sheet at each state change. One sheet stretched across a transformation is how you get a lantern that’s 70% dark in one shot and 40% dark in the next.
Do Runway, Kling, and Seedance handle this differently?
Same underlying idea, three different implementations, and knowing the right field name saves you from re-learning this from scratch on every tool.
- Runway’s Gen-4 References takes up to three reference images per generation, and each one becomes what Runway’s help center calls a distinct visual instruction the model follows through entity-level encoding.
- Kuaishou’s Kling AI shipped a dedicated Multi-Image Reference feature aimed at exactly this problem, and the newer Kling VIDEO 3.0 Omni line splits it further into Element Reference and Video Element Reference, so a prop’s reference can come from a video clip instead of a still image.
- Seedance 2.0 launched February 7, 2026 as ByteDance’s ground-up multimodal rebuild. It accepts up to 12 reference assets in a single generation, nine images plus three video clips plus three audio files, tagged directly in the prompt with an @image1-style syntax so you tell the model exactly which reference applies to which element.
None of these tools ship a button labeled “lock my prop.” They ship a reference mechanism built mainly for faces, and it’s on you to point it at the sword instead.
Runway’s Gen-4 References documentation (help.runwayml.com), one of the reference-image tools this piece covers
The mistake that wastes the most regenerations
Building the character sheet and the prop sheet as one image. It feels efficient. It isn’t.
- If the character is holding the prop in the reference image, every angle change also re-guesses the grip, and the prop shape drifts as a side effect of the hand moving.
- If the prop has reflective or transparent material (metal, glass, gemstones), a single static reference angle won’t carry information about how it should catch light in a different framing, so keep at least one close, evenly lit reference specifically for material.
- If two characters both interact with the same object in a scene, that contact point is where models break fastest. Multi-character contact shots deserve their own test generation before you commit to a full sequence.
How this played out on Lost Garden
Every knight wears a small silver medallion, a detail that matters to the story and has to read the same in a wide establishing shot as it does in a tight close-up. Early on I was still describing it fresh in every prompt: “a silver pendant, engraved, worn on a cord.” Close enough to pass a glance, wrong enough that two shots cut back to back looked like two different objects.
The fix was boring, which is usually the sign a fix is correct: I generated the medallion by itself, three angles, flat lighting, no character attached. Locked that as the reference. Every subsequent shot that needed the medallion pulled that exact reference image in, on top of the character reference, instead of re-describing it. The drift stopped. On my platform the shot list now carries a reference-asset field for exactly this, so a recurring object doesn’t quietly depend on someone remembering to attach the right file three weeks into production.
Do I need a reference sheet for every prop in my film, even background ones?
No. Reserve this for anything that recurs across multiple shots or that the story depends on being recognizable. A prop that appears once in one shot isn’t worth the setup.
Can I reuse a character’s held-prop pose as the prop reference?
Better not to. Generate the prop unheld first, then let the character-plus-prop combination happen inside the actual shot generation, once both references are locked separately.
What if my video tool doesn’t support multiple reference images at all?
Fall back to keeping the exact same descriptive sentence for the prop in every prompt, word for word, and accept a higher rate of regeneration. It’s a worse tool, not a reason to skip the discipline.
Does this apply to 3D or environment assets too, not just handheld props?
The same logic holds for any recurring visual element: a specific building, a vehicle, a piece of furniture that shows up across a sequence. Lock it once, reference it everywhere, and don’t re-describe it from memory.
If your character’s face has been holding steady for weeks and the object in their hand still hasn’t, it’s not a worse model. It’s a reference you haven’t built yet.
