Kandinsky 6.0 Pro Image-to-Video Prompt Recipes: Copy-Paste Prompts for Better Motion

Why prompt structure matters more than prompt length

Image-to-video prompting is not a longer version of text-to-image prompting. When you provide a reference frame, the model already knows what the scene looks like. Your prompt's job is to describe what changes over the clip, not to re-describe what is already visible.

That shift in responsibility is why a 12-word structured prompt usually beats a 120-word paragraph. A long paragraph mixes subject description, motion intent, camera behavior, and style wishes into one undifferentiated blob. The model has no way to tell which clause governs the camera and which governs the subject, so it compromises: the camera drifts a little, the subject moves a little, and the result feels mushy.

A structured prompt separates concerns into layers:

• Subject anchor — a short pointer back to what is in the reference image, confirming what should stay stable.
• Motion cue — the specific physical change you want, with a direction and an intensity.
• Camera phrasing — a dedicated clause that describes only the virtual camera.
• Style constraint — the look, texture, and consistency guardrails.

Keep each layer to one sentence or phrase. The goal is not to be terse for its own sake, but to be parseable: each layer is a variable you can change independently during iteration. A layered prompt is a debugging tool as much as a generation tool.

One more reason structure wins: reproducibility. When a clip turns out well, you need to know why it turned out well so you can reuse it. With a paragraph, the cause is buried. With layers, you can copy the winning layer into the next experiment and vary only the one you are testing.

The practical rule

If a word in your prompt does not describe the reference image's subject, a motion, the camera, or a style constraint, it is probably noise. Cut it, generate again, and compare. More often than not, the shorter structured version holds the frame better and moves more cleanly.

What this guide assumes

Everything below is written for kandinsky-6.0-pro image-to-video through the HiAPI.ai unified interface, where image-to-video calls take a reference image plus a prompt and return a generated clip. The recipes are plain text — no special syntax — so they work in any client that calls the same endpoint.

The anatomy of an image-to-video prompt

Treat the prompt as a form with four slots. Fill them in the same order every time and your experiments stay comparable.

Slot 1: Subject anchor

A brief phrase naming what must remain recognizable. This is not a full description — the image already carries that. It is a stability signal.

Examples: the woman, the ceramic vase, the mountain ridge, the abstract shapes.

Avoid re-specifying details that are visible in the frame. Re-describing clothing, lighting, or color invites the model to "improve" them, which shows up as flicker or identity drift.

Slot 2: Motion cue

One primary motion, stated with a verb, a direction, and an intensity qualifier.

Examples: turns her head slowly to the left, gently sways side to side, drifts steadily upward.

Pick exactly one primary motion per clip. Two competing motions of similar magnitude is the single most common cause of muddy output.

Slot 3: Camera phrasing

A separate clause that starts with camera language: camera, the camera, or shot. Keeping the subject out of this clause is what makes camera behavior predictable.

Examples: camera slowly pushes in, camera holds static, camera pans left at a steady rate.

Slot 4: Style constraint

Guardrails for look and consistency. Short is fine.

Examples: photorealistic, consistent lighting, no flicker, flat illustration style, stable linework, cinematic color grade, shallow depth of field.

Assembled

the woman, turns her head slowly to the left, camera holds static, photorealistic, consistent lighting

That is the whole anatomy. Eighteen words, four slots, one controllable idea. Every recipe in the rest of this guide is this same shape with different content in slots 2 and 3.

Slot discipline

Two habits make the anatomy stick:

• Fixed order. Always subject, motion, camera, style. When you read old prompts back, the order tells you instantly which layer you edited.
• One idea per comma-separated clause. Commas act as soft separators; if a clause contains both a subject action and a camera move, split it.

Motion cue recipes for common subject types

Motion phrasing works differently depending on what is in the frame. These are copy-paste starting points — swap the bracketed parts for your own subject.

People and characters

Human subjects are the most sensitive to over-animation. Small, specific motions read as natural; broad ones read as uncanny.

• [the person], blinks and turns their head slightly toward [direction], camera holds static, natural skin texture
• [the person], smiles subtly while breathing, shoulders rise and fall gently, camera holds static
• [the person], walks slowly toward the camera, steady gait, camera pulls back at the same pace
• [the person], raises one hand to [gesture], other hand stays still, camera holds static

Use subtly, slightly, and gently liberally for faces. They are intensity dampers.

Objects and products

Objects want one clean transformation. Rotation and light movement are the safest.

• [the object], rotates slowly clockwise, camera holds static, product photography lighting, no flicker
• [the object], subtle highlight travels across its surface, object stays still, camera holds static
• [the object], gently floats and bobs in place, soft shadow moves with it, camera holds static
• [the object], lid opens slowly, body remains stationary, camera holds static

Nature and environment

Outdoor scenes tolerate more motion because viewers expect ambient movement — but it should be layered from large to small.

• [the landscape], clouds drift slowly to the right, camera holds static, consistent lighting
• [the landscape], water ripples gently, [foreground element] sways in a light breeze, camera holds static
• [the landscape], fog rolls slowly forward from the background, camera holds static
• [the landscape], sun light shifts from warm to cool across the scene, camera holds static

Abstract and stylized inputs

Abstract frames have no physical logic to fall back on, so name the motion geometry explicitly.

• [the abstract shapes], expand outward from the center at a steady rate, camera holds static
• [the abstract shapes], flow in a continuous horizontal loop, camera holds static, stable palette
• [the abstract shapes], slowly morph into softer forms, camera holds static
• [the abstract shapes], pulse gently in place, rhythm stays even, camera holds static

Reading the pattern

Every recipe above has the same skeleton: verb + direction + intensity. When a clip misbehaves, the fix is almost always in one of those three pieces — wrong verb (rotation vs. translation), missing direction, or missing intensity damper.

Camera phrasing recipes

Camera language only works when it is isolated. These phrasings assume the subject clause already exists and does not mention the camera.

Static framing

• camera holds static, locked-off shot
• camera is completely still, no movement

The explicit no movement phrasing is worth using when a scene is busy; it gives the model a hard constraint to hold against ambient motion in the subject.

Pans

• camera pans slowly left across the scene
• camera pans right at a steady rate
• camera pans horizontally, constant speed

Always include a direction and a rate word (slowly, steady, constant). Without them, pans tend to accelerate or reverse mid-clip.

Pushes and pulls

• camera slowly pushes in toward the subject
• camera pulls back gradually, revealing more of the scene
• camera dolly-in at a slow, even pace

Push and pull are the strongest single camera moves for making a static reference image feel alive. Pair them with an almost-still subject for maximum control.

Orbits and arcs

• camera orbits slowly around the subject, arc of about 30 degrees
• camera arcs to the right, maintaining distance from the subject

Orbiting is the most failure-prone move for image-to-video, because revealing unseen sides requires the model to invent geometry. Keep arcs short and say so. If the clip breaks down, reduce the arc phrase to slight arc or fall back to a push-in.

Tilts and vertical moves

• camera tilts upward slowly
• camera rises vertically, keeping the subject centered

Combining subject and camera motion

When both the subject and the camera move, keep their directions compatible rather than opposing:

the runner, jogs slowly toward the camera, camera pulls back at the same pace, outdoor daylight, consistent lighting

Subject moving forward plus camera retreating reads as a stable tracking shot. Subject moving forward plus camera also pushing in reads as a collision course and usually distorts.

One camera move per clip

Resist stacking pans with pushes with tilts. A single named camera move produces a predictable clip; two competing moves produce a compromise that reads as drift. If you need a compound move for a specific shot, test it alone before combining it with subject motion.

A copy-paste recipe template you can adapt

Here is the full template. Fill in the brackets, delete any optional lines you do not need, and keep the clause order fixed.

[subject anchor], [motion verb] [direction] [intensity], camera [camera move] [rate], [style constraint], [consistency guardrail]

Worked example: product clip

the ceramic vase, rotates slowly clockwise, camera holds static, studio product lighting, consistent shadows, no flicker

Worked example: portrait clip

the woman, turns her head slightly to the right and blinks, camera slowly pushes in, photorealistic, natural skin texture, consistent lighting

Worked example: landscape clip

the mountain ridge, clouds drift slowly to the right, camera pans slowly left, cinematic color grade, consistent lighting

Building your own in four steps

1. Write the subject anchor as a single noun phrase. Do not describe — just name.
2. Choose one motion from the recipes in this guide and edit only the direction and intensity words.
3. Choose one camera phrase and make sure it does not contradict the subject motion.
4. Add a style constraint and one consistency guardrail (consistent lighting, no flicker, stable palette).

Keep a plain-text file of your filled-in templates, one per line, named by intent (portrait-headturn, product-rotate, landscape-clouds). That file becomes your recipe library, and every new experiment starts by copying an existing line rather than writing from scratch.

Versioning the template itself

If you find yourself adding a fifth slot repeatedly — say, a lighting clause — promote it into the template deliberately and update all recipes at once. Ad-hoc additions are how templates rot into paragraphs.

Iterating in controlled steps

Prompt iteration fails when you change several things at once and cannot attribute the result. The fix is procedural, not creative.

Change one variable per run

Generate a baseline, then change exactly one of these per run:

• the motion verb
• the direction
• the intensity qualifier
• the camera move
• the camera rate
• the style constraint

Everything else stays byte-identical. This is the same discipline as a controlled experiment, and it is the only reliable way to learn what a model responds to.

Keep a comparison log

A log can be a plain text file or a spreadsheet. For each run, record:

• recipe name
• the full prompt text
• which single variable changed from the previous run
• a short verdict: keep, reject, or needs another pass

Because prompts are plain text, pasting the full prompt is cheap and makes runs fully reconstructible. Do not summarize prompts in the log — copy them exactly.

Compare against a fixed baseline

The most useful habit is maintaining one canonical recipe per subject type and regenerating it whenever you change anything else in your pipeline. If the baseline still looks right, your pipeline is stable and any differences are attributable to your prompt edit. If the baseline changes, stop prompt work and investigate the pipeline first.

Batch small, review in groups

When testing a direction or intensity word, run a small batch of variations rather than a single clip. Review them side by side and pick the best, then delete the rest from the library. Keeping every variant turns the recipe library into an unreadable archive.

Promote winners, retire losers

When a variation clearly beats the baseline, overwrite the baseline entry in your library and note the date. The library should always contain the current best-known recipe for each intent, not a history of every attempt.

Common failure modes and how to reframe prompts

Most image-to-video problems are prompt problems wearing a settings costume. Before touching any other parameter, try the reframes below.

The clip is over-animated

Symptom: everything moves — subject, background, lighting — and the frame feels noisy.

Reframe: add intensity dampers (slowly, subtly, gently), explicitly freeze the camera (camera holds static, no movement), and add a consistency guardrail (consistent lighting, no flicker). If it persists, remove a secondary motion from the subject clause.

The clip is nearly static

Symptom: the output is essentially the reference image with faint shimmer.

Reframe: make the motion verb stronger and more physical (rotates instead of shifts, walks toward the camera instead of moves). Add a direction if it is missing. If the subject must stay still, add a camera move instead — a slow push-in creates perceived motion without touching the subject.

The subject identity drifts

Symptom: faces, logos, or product shapes warp over the clip.

Reframe: shorten the subject anchor and stop re-describing visible details. Add subject remains unchanged or stable [feature] to the style constraint. Large camera moves that reveal new angles make drift worse — reduce the arc or drop it.

The camera fights the subject

Symptom: the frame feels unstable even though both clauses look reasonable.

Reframe: check directions. Opposing vectors (subject moving left, camera panning right at speed) cause visible conflict. Align them, or reduce one to slowly.

Two motions compete

Symptom: neither motion completes; both look half-hearted.

Reframe: split into two recipes and generate one motion per clip. If you truly need both, order them (first [motion A], then [motion B]) and accept a looser result.

Style drifts toward generic

Symptom: the clip looks smoother and more "AI-ish" than the reference.

Reframe: restate the medium explicitly in the style slot (flat illustration, 35mm film look, product photography), and add stable palette, consistent linework where relevant.

The general rule

If a settings change and a prompt change would both plausibly fix the issue, try the prompt change first — it is faster, free to test, and keeps your pipeline comparable across runs. Reach for settings only when a reframed prompt still fails in the same direction.

Once a recipe survives these reframes, save it. A small, battle-tested library of structured recipes is worth more than any single clever prompt, and it is what makes kandinsky-6.0-pro image-to-video work reproducible across your pipeline.