Kling 4.0 swaps the single prompt for up to 10 keyframes and 15 references per clip

Earlier text-to-video models took one prompt and returned one continuous clip, according to the dev.to explainer. That held up for 5 to 10 second clips ^[7]. Past that, characters changed faces mid-shot and rooms rearranged themselves ^[7]. Kling 4.0 takes up to 10 keyframes instead ^[12]. Spread evenly across a 30-second clip, that is 3 seconds per keyframe ^[17]. The explainer does not describe the request format, so whether keyframes carry timestamps, durations or only an order is unknown.

The explainer's strongest engineering claim is about iteration. It wrote: "If shot three is wrong, you adjust shot three" ^[9]. That holds only if the service can regenerate one segment and leave its neighbors untouched. If the call re-renders the whole clip, a bad shot three sends the caller back through everything.

Omni Reference applies the same idea to consistency. At most 15 of the 50 pooled files are active in one generation, which is 30 percent of the pool ^[18]. Choosing which 15 becomes the caller's job. The explainer frames consistency as an input problem ^[16]. It also hedges: "If the reference system works as advertised, it is a genuine leap." ^[15]

For continuity to transfer to your product, the characters and products it shows have to be coverable by the files you can supply within 15 slots. A product with a large cast or many SKUs per sequence will hit that ceiling before it hits the 10-keyframe one.