A creator spent three generations getting a knight’s face right on Lost Garden, but the medallion around his neck showed three different engravings across those same three shots. AI video models can hold a face steady across a scene, yet an object in a character’s hand is still treated as a sentence, not a subject, and sentences drift. 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, which the model encodes 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 point that feature at the character, while the sword, the medallion, the cup, or the battered lantern usually stay in the text prompt, and a text description is a suggestion, not a lock. The fix is to build a prop reference sheet the same way as a character sheet: generate the prop alone, in image generation, from two or three angles, with nothing holding it, lock the sheet before any video generation starts, attach it as an actual reference asset, keep descriptor language identical in every prompt, and combine the prop with the character at shot time, not before.