I would use the same clear image of the actual product as a subject reference for every clip, so its color, proportions, logo and material have a stable anchor. When one action must continue into the next shot, also use the previous clip's returned last frame as the next clip's first frame, where the chosen model supports those inputs. Chaining only last frames can carry a distorted detail forward; using only the original product image keeps identity steadier but may not connect motion naturally. Test two short clips before committing to the third.

Ask the Buda Agent for a shot sheet: camera size and direction, details that must not change, background and lighting, and the action at each cut. Keep aspect ratio and color treatment consistent. Inspect every generated clip for product labels and silhouette before continuing; in the final edit, cut on an action or use a brief detail shot to bridge minor discontinuities. Generation cannot guarantee an exact product identity across separately rendered clips, so a person should check product-critical frames.

The video generation guide describes returnLastFrame, reference images and first-frame inputs. Are these three clips one continuous action or three separate scenes featuring the same product? The Chinese discussion covers the same approach.