Generating a single impressive clip is easy; producing a sequence where the same subject survives every cut is not. Kling 4.0 is a text, image and reference driven video generator that exposes keyframes and an Omni Reference mechanism for exactly this problem, and the product page states output up to 30 seconds and 4K. This note documents a working order of operations for multi-shot work, plus a short taxonomy of the failures that show up when the inputs are not prepared.
Everything below assumes you already have the visual material. The model does not invent a consistent character for you; it preserves one you hand it.
A sequence prompt written as one paragraph produces drift. The practical unit of work is the shot, not the project. Write the shot list first, in a plain text file, with one line per shot:
s03The reason to fix the wording is that the same noun phrase used in two shots behaves like a soft constraint on identity. Rewriting it between shots, even into a synonym, loosens that constraint and the face or product starts to slide.
Generate or select a single frame that represents the subject well: neutral pose, even light, no motion blur. Use that frame as the image input for the first shot, and keep it as the reference for every later shot.
Practical rules that survived testing:
Where a single prompt leaves too much to chance, define a start frame and an end frame and let the model interpolate. This converts an open-ended prompt into a constrained problem, and it is the main lever for camera moves that must land on a specific composition.
A workable configuration object for one shot looks like this, in pseudocode:
shot = { "id": "s03", "mode": "image_to_video", "start_frame": "refs/hero_front_01.png", "end_frame": "refs/hero_front_01_look_left.png", "reference": ["refs/hero_face_01.png", "refs/product_shell_01.png"], "prompt": "same subject, slow dolly in, soft key light from camera left, no text", "duration_s": 10, }
The fields that matter most are start_frame, reference and the consistency of prompt. Everything else is a preference.
Once reference images carry identity, the prompt should describe only what changes: motion, camera, light direction, atmosphere. Repeating appearance details in the prompt competes with the reference and tends to win, which produces a subject that looks like the description rather than the reference.
A useful sanity check is to read the prompt and ask whether any clause would be true of every shot. If yes, it belongs in the reference, not in the prompt.
Most bad output falls into four buckets, and each has a different fix.
Identity drift. The face or product changes between shots. Cause: prompt re-described the subject, or the reference frame was low quality. Fix: freeze the prompt wording and reuse the same reference set.
Composition drift. The camera move ignores the requested framing. Cause: prompt described motion without a target frame. Fix: add an end frame, or shorten the shot so there is less room to wander.
Artifact accumulation at the end of long clips. The last second degrades. Cause: duration pushed past what the inputs support. Fix: split into two shorter shots and join them in an editor.
Reference bleed. Background or wardrobe elements from the reference appear where they should not. Cause: the reference frame contained clutter. Fix: regenerate the reference with a clean background.
Set a budget before starting: for each shot, allow at most three generations. If the third attempt still fails on identity, stop and fix the reference rather than the prompt. In practice, prompt iteration converges quickly when the reference is good and does not converge at all when it is not.
The same rule applies at the sequence level: if two consecutive shots fail for the same reason, the problem is upstream in the shot list, not in the model settings.
A short sequence built this way takes longer to plan and much less time to regenerate. The model is doing interpolation and preservation; the planning is what keeps the subject recognisable from the first frame to the last.