Blog

Guide

Kling 3.0 Omni Transformation Prompts: Copy-Paste Recipes With Real Outputs

Restyle instructions, subject preservation, 4K notes, and negative prompts you can reuse

Nov 12, 2026 13 min read
Cover image: bone-white sheets stacked with an offset like misregistered overprinting, with a rust-red tissue sheet inserted, set off by crisp studio shadows.
On this page
Kling 3.0 Omni transformation recipe: a locked-off push-in restyled from neutral studio to painterly storybook, subject and pose preserved.

What kling-3.0-omni/transformation actually does

kling-3.0-omni/transformation is a video-to-video endpoint. You give it a source clip plus a text instruction, and it returns a new clip where the visual look has been restyled while the underlying motion and timing come from your input. It is not text-to-video, and it is not a re-shoot. The prompt you write is a restyle instruction aimed at an existing clip, not a description of a scene you are inventing.

That distinction drives everything else in this guide. Because the source clip already carries the subject, the framing, and the motion, your prompt has two jobs at once:

  • Change: state the new visual treatment — medium, palette, lighting, texture.
  • Preserve: state which qualities of the source must survive the restyle — identity, motion path, composition, timing.

If you only write the first half, the model has freedom to reinterpret the second half too, and you get drift: a face that no longer matches, a camera move that no longer lands, a subject that mutates mid-clip. Good transformation prompts read like briefs to a colorist or stylist who can see the source clip — they say make it look like this, and don't touch that.

The mental model: source clip + restyle instruction → restyled clip. Your prompt is the instruction. Treat it as a build artifact you version, not a one-off sentence.

What the prompt cannot do

A restyle prompt is not a storyboard tool. If the source clip does not contain a given element, asking the prompt to add it usually produces artifacts rather than a clean result. Keep your instructions aimed at transforming what is already there. Similarly, timing is inherited from the input — you are not directing new motion, you are asking for a new look over existing motion.

A quick pipeline sketch: source clip → transformation endpoint with a prompt → restyled output clip. Everything in the recipes below is about what goes into the middle step.

Anatomy of a working restyle prompt

Length is not the lever; structure is. A working restyle prompt has four ordered slots, and each slot answers a different question. Keep the order stable so that revisions are comparable.

Slot 1 — Change: what the output should look like

Lead with the target look. Be concrete about medium and treatment: "stylized 3D animation with soft clay shading" beats "make it cool." Name the palette, the lighting logic, and the surface texture you want. This is the intent of the restyle and belongs first because it frames how everything after it is read.

Slot 2 — Preserve: what must survive

Immediately after the change, lock down the invariants. State that the subject identity, clothing, framing, camera motion, and timing are kept from the source. Write this explicitly even when you assume it is obvious — in practice, explicitness is what reduces drift. Preservation is two separate constraints (subject and style), and the next recipes section has wording for each.

Slot 3 — Style: the target treatment in detail

This is where you refine the look from slot 1 with the specific vocabulary of the style you want — rendering characteristics, edge treatment, grain, contrast behavior. Keep it descriptive and visual. This slot should not introduce new subjects or new camera moves; it only sharpens the look.

Slot 4 — Constraints: what to avoid and how far to push

Close with boundaries: what the restyle must not become, and negative prompts for common failure modes. Constraints go last so they act as a final filter on the interpretation of the earlier slots.

The skeleton

Restyle the source clip into <target look>.
Keep <subject identity, clothing, framing, camera motion, timing> unchanged from the source.
Render with <style detail: palette, lighting, texture, edge treatment>.
Avoid <drift modes>; do not <specific failure>.

Why order matters

Wording order changes how instructions interact. Change-first gives the model a target before it reads the constraints. Preserve-second narrows the space of valid outputs early. Constraints-last filters the result without rewriting the intent. When a prompt misbehaves, reorder before you lengthen — moving the preserve clause earlier is often a faster fix than adding more adjectives.

Copy-paste recipes with before/after clips

Each recipe below is written for the same source clip: a short single-subject shot with steady camera motion. To compare fairly, use the identical source clip for every variant and watch the paired before/after clip for the recipe. Swap only the bracketed portions when adapting to your own footage.

Recipe 1 — Stylized 3D restyle

Restyle the source clip into a stylized 3D animated look with soft clay-like surfaces.
Keep the subject's identity, clothing, and facial features exactly as in the source, and keep the original framing and camera motion unchanged.
Render with a warm palette, soft ambient lighting, subtle surface grain, and clean rounded edges.
Avoid photoreal skin texture; do not alter the subject's proportions or the motion timing.

Before/after: the clip keeps the same shot and motion, with the subject re-rendered as a readable clay-style figure; identity and framing stay recognizable.

Recipe 2 — Graphic ink and halftone

Restyle the source clip into a high-contrast graphic ink illustration with halftone shading.
Preserve the subject's silhouette, pose sequence, and the original camera movement; keep frame timing identical to the source.
Render in a restrained two-tone palette with visible line work and paper-like texture.
Avoid muddy midtones, photorealistic detail, and any changes to the subject's face or body proportions.

Before/after: motion reads the same; the surface treatment becomes flat and graphic, with the silhouette preserved.

Recipe 3 — Vintage film treatment

Restyle the source clip with a vintage analog film look: soft contrast, gentle grain, and warm highlights.
Keep the subject identity, wardrobe, composition, and all camera and subject motion unchanged from the source.
Render with mild halation, slightly lifted blacks, and a stable frame without jitter or gate weave.
Avoid modern digital sharpness and do not introduce flicker or color shifts between frames.

Before/after: the same shot gains a filmic surface without the instability that vintage looks often introduce.

Recipe 4 — Clean toon shading

Restyle the source clip into clean cel-shaded toon animation with flat color fills and crisp outlines.
Preserve the subject's identifying features, clothing design, framing, and motion exactly as in the source.
Render with a limited palette, consistent line weight, and even lighting that stays stable across the clip.
Avoid sketchy or inconsistent outlines; do not change the subject's proportions or the original timing.

Before/after: identity and motion carry through; the treatment becomes flat and illustrative with stable line work.

How to read the before/after pairs

Judge three things on every pair: does the subject still read as the same subject; does the motion and timing still match the source; does the style stay consistent across the whole clip rather than only in the first frames. Those are the three failure axes the next two sections help you control.

Preserving subject and style while restyling

Subject preservation and style preservation are separate constraints, and they fail in different ways. Subject preservation keeps who is in the clip; style preservation keeps how the clip is shot and cut. Write both explicitly, and test them separately.

Subject preservation patterns

Use identity-anchored phrasing rather than a vague "keep it the same":

  • "Keep the subject's identity, facial features, and clothing exactly as in the source."
  • "Preserve the subject's silhouette and body proportions; do not restyle the subject's face."
  • "Hold the subject's pose sequence and gestures identical to the input."

Specific nouns beat adjectives. "Facial features and clothing" gives the model a checklist; "keep it the same" does not.

Style preservation patterns

Style here means the structural properties of the shot that carry continuity — framing, camera motion, timing, lighting direction:

  • "Keep the original framing, camera motion, and shot length unchanged."
  • "Maintain the source lighting direction and shadow placement."
  • "Preserve the motion timing; do not retime or reorder the action."

The trap: readers assume the transformation endpoint preserves these automatically. State them anyway. When style constraints are omitted, the model may treat changes to framing or motion as part of the restyle.

Separating the two in the prompt

Put subject preservation immediately after the change clause, and style preservation directly after it. Keeping them adjacent makes it easy to see which constraint a revision touched. If you edit one, edit only one per revision so your before/after comparison isolates the effect.

When both conflict

Sometimes a heavy style transfer and strict subject preservation pull against each other — a flat graphic style naturally flattens facial detail. Decide which constraint wins and say so: "Prioritize subject identity over style fidelity if they conflict." Stating the priority produces more predictable output than leaving the model to guess.

Negative prompts and drift control

Negative prompts are the fastest lever when output drifts. Drift is any unwanted change from the source or from the intended style, and it usually falls into a few recognizable categories. Write one negative clause per category you actually see.

Common drift categories and wording

  • Identity drift: "no changes to the subject's face, body proportions, or clothing."
  • Motion drift: "no retiming, no reordered action, no added camera movement."
  • Style instability: "no flicker, no frame-to-frame flicker, no inconsistent outlines."
  • Style overshoot: "no extra rendering effects beyond the specified treatment; avoid heavy glow or bloom."
  • Artifacts: "avoid warped edges, smeared textures, and doubled limbs."

Keep negatives specific to observed failures. A long generic negative list dilutes the constraint; two or three targeted ones usually do more.

Reacting to drift without guessing parameters

The temptation when output drifts is to hunt for a knob. Order your fixes from cheapest to most invasive:

  1. Rewrite, don't add. Tighten the change or preserve clause before appending anything new.
  2. Reorder. Move the violated constraint earlier in the prompt.
  3. Strengthen the negative. Add a precise negative for the exact artifact you see, phrased as the failure rather than the goal.
  4. Narrow the style. If the style overshoots, remove style detail from slot 3 rather than adding more constraints.

Only after the prompt has been reorganized should you consider input-side changes — for instance using a source clip with more stable framing or a clearer single subject. Do not invent parameter values; if the effect cannot be achieved through prompt wording, adjust the source clip or the recipe.

Wording-order lever

When drift is intermittent — some frames good, some bad — the fix is usually placement, not content. Put the constraint governing the unstable element right after the change clause so it binds to the target look, and repeat the key preservation phrase in the constraints slot so it survives to the end of the prompt.

4K output notes for practical runs

Targeting high-resolution output has its own set of things to watch. These are checklist items, not guarantees — verify each on your own clip rather than assuming.

Validate the recipe at low resolution first

Prompt behavior should be checked on a short clip at a modest resolution before you commit to a high-resolution run. Restyle judgment is about look and preservation, both of which are visible without full resolution. Only once a recipe is settled should you run it at the resolution you need for delivery.

Watch detail stability across the clip

High-resolution output shows fine detail more clearly, which cuts both ways: your style treatment becomes more legible, but flicker, grain inconsistency, and edge artifacts become more visible too. Check a few non-adjacent frames in the before/after pair for consistency, not just a single flattering frame.

Keep the source clip clean

Detail that survives the restyle comes from the source. A clip with motion blur, compression noise, or an ambiguous subject gives the model more to reinterpret, and high-resolution output renders that noise faithfully. A clean, well-lit, single-subject source is the most reliable input for any resolution.

Lock style wording when you raise resolution

If you change the recipe while also changing resolution, you cannot tell which change caused a difference. Freeze the prompt text, run the resolution change, then judge. Treat resolution as a separate variable in your evaluation.

Plan for longer runs

Higher-resolution transformation of a longer clip takes more time and is best scheduled as a batch rather than interactively. Prepare the source, the frozen prompt, and the comparison expectation before launching so that a run is one clean step.

Checklist before a 4K run

  • Recipe validated on a short, low-resolution clip
  • Source clip clean and single-subject
  • Prompt text frozen and versioned
  • Before/after judged on non-adjacent frames
  • Plan for the batch run and storage of the output

Turning recipes into an evaluation loop

A prompt recipe is only useful if you can tell whether a revision made things better. That means treating the source clip and the prompt as a fixed pair and changing one variable at a time.

Fix the source clip

Pick one representative source clip and keep it constant across all experiments. Every recipe in this guide was authored against the same footage for exactly this reason: a stable source makes before/after comparisons trustworthy, and a swap changes more than the prompt does.

Version prompts like code

Keep each prompt revision as a discrete, named version with the change noted — for example, "v3: moved subject preservation before style detail." One revision changes one thing: a reorder, a strengthened negative, or a narrowed style slot. If you edit two things at once and the output improves, you have learned nothing reusable.

Score against fixed axes

For each version, score the before/after pair on the axes that matter for production: subject identity preserved, motion and timing preserved, style consistency across the clip, and artifact presence. Keep the scores next to the prompt version so the rationale survives the run.

Shortlist, then scale

Run every candidate recipe on the short test clip first. Shortlist the versions that pass all four axes, then run those on longer or higher-resolution inputs. Scaling a prompt that already passed the short-clip gate costs far less than discovering drift in a production batch.

Wire it in

When a version wins, freeze it as the production recipe, attach its before/after reference clip to the asset, and route it through the kling-3.0-omni/transformation endpoint with the same ordered structure you validated. Keep the losing versions recorded — they document why the chosen wording exists, which is exactly what the next person editing the prompt will need.

The loop is simple: one source clip, one change per revision, fixed scoring, scale only after the short clip passes.

Related articles

View all articles →

Start building with persistent AI artifacts

Get your API key and call image, video and audio models through one task-based API.

Read the API docs