Lore
Start/end frame anchoring for label-safe product animation
Instead of asking a video model to invent a whole product clip from one reference, generate two clean still frames first, a start frame and an end frame, and have the video model (Kling 3.0) animate between them.
Why it works
- Both endpoints are stills made by an image model from the real product image, so the product, its label and its small text are correct at the start and at the end.
- The video model only has to solve the middle of the clip. That leaves it "less room to hallucinate during the animation".
- Because the end frame is built from the original product image, every clip lands on a correct label. Any warping is confined to mid-clip, where movement hides it (see motion-masks-mid-clip-text-warping).
Input
The only required input is a simple reference image of the product, for example a white-background packshot. No multi-angle shoot is needed.
Stress-testing the claim
The author picked label-heavy products on purpose: the Ordinary (small text) and Liquid Death (complicated front graphic). They are where AI is weakest, so a clean result there says more than an easy product would. Judge a label-safety workflow on its hardest products.
Relationship to other approaches