Guide
Kling 3.0 Omni Image-to-Video Prompt Recipes: Copy-Paste Templates for Developers
First/last-frame structure, camera motion wording, native audio cues, and a stability negative list — with real generated clips
On this page
Why Prompt Structure Beats Prompt Length
Most image-to-video prompts fail not because they are too short, but because they are unordered. A single paragraph that mixes subject description, camera intent, audio mood, and repair instructions gives the model no deterministic map from text to motion. When a clip comes out wrong, you can't tell which clause caused it — so you can't fix it, only reroll.
The practical fix is section discipline. Treat the prompt as a fixed five-block contract, always in the same order:
- Frame anchor — which image is the start, which is the end, and how strict that binding is.
- Subject — who or what is in frame, with the identity details that must survive the transition.
- Motion — what physically changes between the frames, in plain temporal terms.
- Camera — how the viewpoint moves (or explicitly holds still).
- Audio + negatives — native sound cues, then the failure modes to suppress.
This order mirrors how the model resolves constraints: it locks the frames first, then decides what moves, then decides where the viewer stands, then fills in non-visual output. Keeping audio and negatives last means visual reasoning isn't perturbed by instructions about sound or repair.
For engineering teams the payoff is reproducibility. A stable skeleton means a diff between two prompt versions is readable: you changed the camera line, not the whole blob. It also means prompts become parameters rather than prose — you can template, version, and A/B them in a pipeline, which is exactly what the later sections build toward.
Length is a symptom, not the lever. A 60-word prompt in the fixed order outperforms a 200-word paragraph that front-loads adjectives. Start by compressing every prompt you already have into these five slots, and delete anything that doesn't fit a slot.
The Base Recipe: A Copy-Paste Prompt Skeleton
Here is the canonical skeleton. Paste it, fill the brackets, delete slots you genuinely don't need — but keep the order.
[FRAME ANCHOR]
Start frame: the provided image, preserved exactly.
End frame: [describe end state, or "none — open-ended continuation"].
Transition: [smooth continuous | single cut-free motion | holds start pose].
[SUBJECT]
[Subject identity + 2-3 invariant details that must not change].
[MOTION]
[What moves, direction, and how far — one sentence per moving element].
[CAMERA]
[Static locked-off shot | slow push-in | left-to-right pan | gentle orbit].
Pace: [slow and even | builds to a faster finish].
[AUDIO]
Ambience: [environment sound].
Music: [feel, not genre-name dropping].
Sync: [what the audio should land on].
[NEGATIVE]
flicker, morphing, warping text, extra limbs, identity drift
A filled example — a product shot of a ceramic cup on a windowsill:
[FRAME ANCHOR]
Start frame: the provided image, preserved exactly.
End frame: same cup, steam gone, light shifted warmer.
Transition: smooth continuous motion, no cuts.
[SUBJECT]
Matte white ceramic cup, thin dark rim, wooden windowsill.
[MOTION]
Steam thins and disappears over the shot; curtain edge drifts slightly.
[CAMERA]
Static locked-off shot, very slow push-in.
Pace: slow and even.
[AUDIO]
Ambience: quiet room tone, faint birds outside.
Music: soft, sparse, low volume.
Sync: birds chirp as steam vanishes.
[NEGATIVE]
flicker, morphing, warping text, extra limbs, identity drift
Run through kling-3.0-omni/image-to-video, this produces a clip where the cup stays a cup, the change is legible, and the audio supports rather than competes. The reason it works is that every clause is doing one job: the anchor constrains geometry, the subject guardrails identity, motion is the only thing allowed to change, camera is declared rather than implied, and negatives are handled as a separate gate.
Two habits to keep: write motion in time order ("thins, then disappears"), and never let the camera and the subject argue — if the subject walks left, don't also orbit left at speed unless you want the frame to feel unmoored.
Slot discipline
Each slot should contain only its own concern. If you catch yourself writing "the camera slowly pushes in while the cup's steam morphs into a bird," you've merged camera and motion into one clause and lost the ability to tune either independently. Split it.
First-Frame and Last-Frame Recipes
Image-to-video with two frames is an interpolation problem, not a generation problem. The model's job is to find a believable path between two fixed states. Prompts fail here when they describe the scene instead of the transition — the model then reinterprets the content and the endpoint drifts.
Recipe: strict endpoint, minimal invention
Use this when both frames exist and you want the model to connect them literally.
Start frame: provided image, preserved exactly.
End frame: provided image, preserved exactly.
The clip must begin on the start frame and end on the end frame.
Do not introduce new subjects or objects. Only intermediate motion.
Transition: continuous, no cuts, no scene change.
Motion: [the single physical change that carries start to end].
Camera: static locked-off shot.
The key sentence is only intermediate motion. It tells the model the endpoints are constraints, not suggestions. Declaring a static camera removes another degree of freedom, which makes the interpolation far more stable.
Recipe: open-ended continuation
Use this when you only have a start frame and want natural extension.
Start frame: provided image, preserved exactly.
End frame: none — open-ended continuation of the same scene.
Transition: continuous motion, same location, same lighting.
Motion: [primary action], resolving within the clip length.
Camera: [declared move].
Negative: scene change, new location, new characters.
Describing the end state without an end image
If there is no last frame but you want a directional outcome, describe the end state in appearance terms rather than events: "liquid settled and still," "door fully open," "steam gone." Appearance states are visually verifiable; event descriptions ("the character decides to leave") invite the model to invent staging.
Two rules that prevent most endpoint failures:
- State the start frame's permanence explicitly. "Preserved exactly" is stronger than "use the image as a reference."
- Keep the transition count at one. If two things must change, let one lead and the other follow, written in sequence. Competing transitions read as a scene change and the model will regenerate rather than interpolate.
Camera Motion and Pacing Wording
Camera language is where vague prompts most often collapse into random drift. The fix is to name the move, name the target, and name the speed — three words of vocabulary per shot.
Moves that map cleanly
- Push-in — "slow push-in toward [subject], centered." Predictable, good for product and portrait reveals.
- Pull-back — "slow pull-back revealing [context]." Use when the start frame is a detail.
- Pan — "left-to-right pan across [scene], constant speed." Always give direction and constancy.
- Tilt — "gentle tilt up from [lower element] to [upper element]."
- Orbit — "slow orbit around [subject], subject stays centered." Orbit is the most failure-prone; the centered clause keeps the subject from sliding out of frame.
- Static — "static locked-off shot, no camera movement." Say it even when you think it's obvious.
Pacing without numbers you can't verify
You don't need to guess frame counts. Describe pacing in adverbs that map to beats: slow and even, steady, builds to a faster finish, settles in the final moment. These translate into acceleration and deceleration that fit short-form beats without you committing to a timing you can't measure.
For short-form rhythm, the reliable pattern is one camera move per clip. A push-in that becomes a pan that becomes an orbit reads as a demo reel, not a shot. If you need a second move, sequence it: "first a slow push-in, then a brief hold."
Keeping camera and subject independent
Separate the lines in the prompt, always:
Camera: slow push-in toward the subject.
Pace: slow and even.
Motion: the subject turns slightly toward the camera.
Three lines, three concerns. When you merge them — "the camera pushes in as the subject turns" — you lose the ability to change one without disturbing the other, and small edits produce large unpredictable swings.
Finally, declare the shot size when it matters: "medium shot," "close-up," "wide establishing shot." Shot size is a cheap constraint that prevents the model from re-framing mid-clip.
Native Audio Prompting
kling-3.0-omni/image-to-video produces native audio alongside the visuals. The most common mistake is mixing audio description into motion sentences, which causes the audio cue to be read as a visual instruction. Keep audio in its own block, written in sound terms only.
Three slots, always in this order
Ambience: [continuous environment sound].
Music: [feel and density, not track names].
Sync: [what the audio should land on].
Ambience is the bed: room tone, wind, street hum, water. Describe it as continuous and low-level unless you want it foregrounded.
Music should be described by feel and density — "sparse and soft," "steady low pulse," "bright and light." Naming genres is weaker than describing texture because the model has to map the phrase to a sound profile; adjectives map more consistently than labels.
Sync is the timing bridge. Name a visual event and the sound that should land on it: "a soft click as the lid closes," "birds as the steam vanishes." One or two sync points per clip is plenty. More than that and the audio becomes busy and the visuals get pulled toward matching every cue.
Cues that work
- Onset: "sound begins as the motion starts" — useful for cuts into a clip.
- Fade: "ambience fades out at the end" — prevents abrupt truncation.
- Emphasis: "a single soft impact when [object] lands."
- Restraint: "no music, ambience only" — a legitimate and often better choice.
Mixing without clashing
If ambience and music both exist, declare their relationship once: "ambience low, music underneath." This costs one clause and prevents the two beds from fighting.
One warning: don't describe audio in onomatopoeia or sound-effect names. "Whoosh" and "ding" are ambiguous across contexts. Describe the source and the character instead — "a short soft puff of air as the cloth lifts." The audio block should read like a sound designer's note, not a comic strip.
Stability Negative List
Negatives are a repair layer, not a creative tool. Keep them short, targeted, and copy-pasteable. Here is the base list, then the failure mode each term addresses.
flicker, morphing, warping text, extra limbs, identity drift,
scene change, jitter, frame tearing, duplicated objects
Failure mode → term
- Luminance flicker across frames →
flicker - Objects melting or blending between frames →
morphing - Signage and captions distorting →
warping text - Hands and limbs multiplying or collapsing →
extra limbs - Face or product identity changing mid-clip →
identity drift - Background regenerating as a new location →
scene change - Micro-shake in otherwise static shots →
jitter - Edge artifacts at frame boundaries →
frame tearing - Repeated objects appearing from motion blur →
duplicated objects
Conditional negatives
Add these only when the failure actually occurs — a long negative list dilutes each term's effect:
- For text-heavy frames:
illegible text, scrambled letters - For product shots:
shape distortion, color shift - For portraits:
asymmetric face, warped eyes - For slow shots:
sudden motion, pop-in
Using the list as a checklist
Treat this as a pre-flight check, not a prompt accessory. Before you run a batch, ask which failure modes your scene is exposed to. A static product shot needs flicker and shape-distortion guards; a moving subject needs identity drift and limb terms; a text-bearing frame needs the text terms. Everything else gets the base list.
When a clip fails, change one thing — add a single negative, or tighten one motion line — then rerun. Batch-editing five negatives at once tells you nothing about which one helped.
Adapting Recipes Into an API Pipeline
A prompt that lives in a notebook doesn't scale. The last step is turning the skeleton into a templated string with variables, so recipes can be versioned, reviewed, and batch-run through the HiAPI API.
Template the skeleton
PROMPT = """[FRAME ANCHOR]
Start frame: the provided image, preserved exactly.
End frame: {end_frame}.
Transition: {transition}.
[SUBJECT]
{subject}
[MOTION]
{motion}
[CAMERA]
{camera}
Pace: {pace}.
[AUDIO]
Ambience: {ambience}.
Music: {music}.
Sync: {sync}.
[NEGATIVE]
{negatives}
"""
Every variable is a slot from the recipe. The template itself never changes — only the variables do — which means a prompt diff is a diff of one line, and a regression is traceable to a specific slot.
Keep recipes as data
Store each recipe as a dict with a name and a version:
recipes = {
"product-static-v2": {
"end_frame": "none — open-ended continuation",
"subject": "matte white ceramic cup, thin dark rim",
"motion": "steam thins and disappears",
"camera": "static locked-off shot, very slow push-in",
"pace": "slow and even",
"ambience": "quiet room tone, faint birds",
"music": "soft, sparse, low volume",
"sync": "birds chirp as steam vanishes",
"negatives": "flicker, morphing, warping text, extra limbs, identity drift",
}
}
Version names, not just numbers — product-static-v2 tells you what changed; v2 doesn't. When a new failure appears, add a new version rather than editing in place, so old outputs stay reproducible.
Batch discipline
- Render one variable at a time when tuning. Batch only after a recipe is stable.
- Log the filled prompt string with every output. Without it you can't attribute a bad clip.
- Keep a shared negative base and let recipes extend it, never replace it.
Where the API fits
The recipe structure is model-facing, so it travels intact to kling-3.0-omni/image-to-video on HiAPI: you send the frame inputs plus the rendered prompt string, and the section order you enforced in the template is the section order the model sees. That's the whole point of the skeleton — it is simultaneously human-readable documentation and the exact string in production.
Start with one recipe, version it, and let the variable names do the talking. For the endpoint details and parameter names, see the HiAPI API docs.
Related articles
View all articles →-
Guide
gpt-6.2-sol Getting Started Guide: Capability Boundaries, API Integration, and Token Pricing Essentials
Integration decisions and practical essentials for engineering teams
-
Guide
gpt-6.2-astra Getting Started Guide: Text Model Capability Boundaries, Integration Essentials, and Cost Control
-
Guide
Tripo H3.1 Getting Started Guide: Choosing Among Three Input Modes, Parameters, and Pricing
A practical guide for developers on 3D generation integration decisions and cost control