Blog

Guide

wan2.6-flash Image-to-Video Prompt Recipes: Copy-Paste Prompts That Work

First-frame descriptions, camera and pacing phrasing, native-audio cues, 1080P fast-tier trade-offs, and a negative-prompt checklist

Nov 20, 2026 14 min read
Flat cover art with dark ink background and interlocking celadon and slate liquid ribbons folding over each other; guide article cover.
On this page

Why Image-to-Video Prompts Need a Recipe, Not a Paragraph

Most people write their first image-to-video prompt as a paragraph — a dense description of everything in the frame and everything they hope will happen. On a text-to-video model that makes sense, because the model has nothing else to go on. On image-to-video it works against you.

The input image already locks in composition: where the subject sits, what the lighting is, which objects are in the background, the color palette, the lens character. If you re-describe all of that in prose, you spend most of your prompt budget on things the model can already see, and you leave very little room to specify the things it cannot: how the subject moves, how the camera behaves over time, how fast or slow the moment should feel, and what the scene should sound like.

Think of the still as a starting state and the prompt as a set of instructions layered on top of it. Each instruction answers a different question:

  • What must stay exactly as it is?
  • What moves, and in which direction?
  • How does the camera move relative to the subject?
  • What is the tempo across the clip?
  • What do we hear?
  • What must never appear?

Those six questions map naturally onto a small number of prompt slots. A recipe is just a fixed order for filling those slots so that every shot you generate is described the same way. That consistency is what makes outputs comparable — and comparison is what lets you improve a prompt instead of guessing at it.

A paragraph invites the model to blend everything together with equal weight. A recipe lets you point at one slot, change one line, and know what caused the difference. The rest of this guide is a set of copy-paste recipes for those slots, plus the practical trade-offs you will hit on a fast 1080P tier.

The still fixes the frame; the prompt should steer its branching streams of motion, camera, pacing and audio.

The Five-Slot Prompt Skeleton

Every recipe in this article is a filling of the same skeleton. Write it in this order, every time:

[1] FIRST FRAME: what is in the shot and what must stay fixed
[2] CAMERA: one move, direction, and speed
[3] PACING: tempo and how the motion evolves over the clip
[4] AUDIO: ambient bed + any event sound, tied to visual beats
[5] CONSTRAINTS: short list of what to suppress

Why the order matters

Put first-frame anchors first, because the model's earliest decisions are about preserving identity and layout. Camera goes second because it is the single largest source of motion in most shots — if the camera move is wrong, nothing else can rescue the clip. Pacing comes third so the tempo instruction can shape the move you just described rather than fight it. Audio fourth, because it should follow the visual beats you already defined. Constraints last, as a compact tail.

A filled example

FIRST FRAME: keep the man's face, jacket and the wet street reflections unchanged.
CAMERA: slow push in, straight toward the subject, no rotation.
PACING: steady, unhurried; motion completes by the end of the clip.
AUDIO: light rain ambience; footsteps are soft and continuous.
CONSTRAINTS: no flicker, no face morphing, no camera shake.

Keeping the skeleton short

Each slot should be one to three short sentences. If a slot grows past that, the extra text is almost always aesthetic adjectives that do not steer motion. Cut them and put anything genuinely important into the constraints slot as a negative.

Two rules for filling it

  1. One camera move per shot. Two moves in one clip is the most common cause of jittery, unreadable motion.
  2. Concrete verbs over adjectives. "The flag lifts and curls" beats "a dramatic, cinematic atmosphere."

The skeleton is the same whether you are animating a portrait, a product still, or a landscape. Only the contents of each slot change — which is exactly what the following recipes cover.

The five-slot skeleton: five stacked slots you fill top to bottom for any shot.

Describing the First Frame So Motion Has an Anchor

The first-frame slot is not a caption. Its job is to tell the model which pixels are load-bearing so it does not reinvent them.

Name the anchors, not the scene

Instead of describing the whole image, list the elements that must survive the clip:

FIRST FRAME: preserve the woman's face and hair style, her grey coat, the
window frame behind her, and the warm lamp on the left side of the frame.

You are not telling the model what the image is. You are telling it which parts are identity-defining: faces, clothing, logos, text, distinctive props, and the main light source direction.

Separate fixed from moving

The most useful sentence in this slot is the one that splits the world in two:

FIRST FRAME: the product bottle and the table stay fixed; only the
background curtain and the light on the bottle surface change.

This single pattern suppresses most unwanted global motion, because the model now knows where motion is allowed to live.

Anchor the starting pose

For people and animals, state the pose at the first frame so the motion has a departure point:

FIRST FRAME: subject is seated, hands resting on the table, head slightly
turned to the left; keep that starting pose at the top of the clip.

Anchor the light

Lighting direction is the fastest thing to drift. One clause keeps it stable:

FIRST FRAME: key light stays from the upper right, shadows unchanged.

What to leave out

  • Do not repeat color grading, film stock, or lens details the image already shows.
  • Do not describe objects that are not in the frame and that you do not want to appear.
  • Do not restate the subject's appearance in flowery terms; a short anchor list is stronger than a paragraph of description.

Treat the first-frame slot as a fence. Everything you name inside it is protected; everything else is free to move — and you get to say how in the next slot.

Fixed anchors on both panels stay identical; only the highlighted current shifts between frames.

Camera and Pacing Phrasing That Actually Moves the Shot

Camera wording does more to change a clip than any other slot, and it is where vague style adjectives do the most damage. "Cinematic camera work" tells the model nothing. "Slow push in, no rotation" tells it everything.

Use a small vocabulary of camera moves

Keep to a short, concrete list and pick exactly one per shot:

  • push in — camera moves toward the subject
  • pull out — camera moves away
  • pan left / pan right — rotate horizontally in place
  • tilt up / tilt down — rotate vertically in place
  • track left / track right — move sideways alongside the subject
  • orbit — circle around the subject
  • static — no camera movement at all; motion comes only from the scene

Always add speed: slow, steady, or fast. And always add the negative case you care about: "no rotation," "no handheld shake," "no zoom."

CAMERA: slow push in, straight on, no rotation, no shake.

Pair one move with one subject action

A clean shot usually has exactly one camera intention and one in-scene action:

CAMERA: slow track right, keeping the subject centered.
PACING: subject turns her head to follow the camera over the full clip,
reaching the final pose in the last third.

Write pacing as a timeline, not an adverb

"Slowly" is weak because it has no reference point. Anchor the motion to the clip's duration:

  • "motion begins at the top of the clip and completes by the end"
  • "hold the first pose for the opening beat, then move"
  • "the action resolves in the final third"

This phrasing keeps the model from rushing the motion in the first second and then freezing, which is a very common failure on short clips.

A pacing recipe for each intent

# calm product shot
PACING: very slow, continuous, no acceleration.

# reveal
PACING: hold, then a single smooth move that completes near the end.

# action beat
PACING: immediate motion from the first frame, constant speed, no pause.

What to avoid

  • Two camera moves in one clip.
  • Stacked adjectives ("epic sweeping dramatic camera") with no direction or speed.
  • Pacing described only as an emotion ("tense") with no timing reference.

If you are ever unsure, choose the simpler move. A static shot with a clear subject action is more usable than an ambitious multi-move that jitters.

Adding Native Audio Cues to the Prompt

When the model supports native audio, the audio slot is not decoration — it is a second timeline that should line up with the visual beats you already wrote.

Split ambient from event

Every audio description has two layers:

AUDIO:
  ambient: steady rain on pavement, distant traffic hum
  event: soft footsteps that begin with the subject's first step and
         continue to the end of the clip

The ambient bed is continuous; the event sounds are timed. Writing them separately keeps you from burying a timed cue inside a wall of atmosphere.

Tie sound to visual beats

Reference the pacing slot so audio and motion share a clock:

PACING: subject lifts the cup over the first half of the clip.
AUDIO: ambient cafe murmur; a single ceramic clink lands as the cup
       settles on the saucer in the second half.

One or two timed events is plenty for a short clip. More than that and the audio becomes busy and hard to describe reliably.

Keep the vocabulary physical

Describe sound the way you would describe an object: source, material, and distance.

  • Good: "dry footsteps on gravel, close to the camera"
  • Weak: "dramatic sound design"
  • Good: "low wind, far away, no voices"
  • Weak: "cinematic ambience"

What to keep out of the audio slot

  • Dialogue and lyrics. Speech is a separate, harder problem; describing spoken lines in the audio slot is unreliable. If you need a silent shot, say so explicitly: "ambient only, no voices."
  • Music genre labels with no behavior. "Lo-fi beat" is less useful than "steady low pulse, no vocals."
  • Sound with no visible source. If you ask for a sound whose cause is not in the frame or in the action, the model has no anchor for when to place it.

A safe default

If the shot does not need audio detail, a single ambient line is better than nothing:

AUDIO: quiet room tone only, no voices, no music.

That line prevents random speech or musical stings and keeps the clip usable while you iterate on the visual slots first.

1080P Fast Tier: When Speed Is Worth the Trade-Off

A fast 1080P tier exists so you can iterate. Treat it as a dial you set per shot, not a project-wide setting.

The trade-off in plain terms

Higher output resolution gives you cleaner detail, more stable fine texture, and better retention of small features — faces at a distance, text, thin structures. Faster tiers trade some of that fidelity and stability for turnaround. The right choice depends on what the shot is for at that moment.

Use the fast tier when

  • You are still searching. Composition, camera move, and pacing decisions are visible even when fine detail is not. Do all of your prompt iteration here.
  • The subject is large and simple. A close portrait or a product on a plain background has little fine detail to lose.
  • The shot is a placeholder. Editorial timing, storyboard animatics, and cut-tests do not need final fidelity.
  • You need many variants. When you are comparing ten candidate motions, speed matters more than polish.

Spend more when

  • Fine texture carries the shot. Fabric weave, hair strands, foliage, water spray, and dense backgrounds are where fidelity differences show first.
  • Text or logos are visible. Small lettering degrades quickly.
  • The clip is a final deliverable. Anything that ships to an audience deserves the higher-fidelity pass.
  • You already locked the prompt. Once the recipe is stable, re-run the winning prompt at the higher setting instead of re-running everything.

A practical workflow

  1. Write the five-slot recipe.
  2. Iterate on the fast tier until camera, pacing, and audio behave.
  3. Change one variable at a time.
  4. Re-run the locked recipe at the higher-fidelity setting for the final.

The key discipline is not to mix the two activities. Do not polish fidelity while the camera move is still wrong, and do not keep re-rolling composition on an expensive setting.

You can try these recipes directly on the wan2.6-flash image-to-video model page, and check current tier details on the pricing panel there, since billing and tier availability live in the site's own pricing view rather than in a fixed number inside a guide.

The Negative Checklist

Negatives are cheap insurance, but a long negative list becomes noise. Keep a short standing checklist and append only the failure you actually saw.

The standing checklist

Start every recipe with this tail:

CONSTRAINTS: no flicker, no morphing of faces or hands, no warping of
straight edges, no unwanted camera drift, no text or watermark artifacts.

That covers the most common failures:

  • Flicker — brightness or texture pulsing frame to frame.
  • Morphing — faces, hands, and small objects blending or dissolving.
  • Edge warping — straight lines (doorframes, table edges, product silhouettes) bending.
  • Camera drift — the camera creeping when you asked for static.
  • Text artifacts — garbled lettering where the source image had clean text.

Add negatives from evidence

Do not pre-emptively write twenty negatives. When a run fails in a specific way, add exactly one line describing it:

CONSTRAINTS: no flicker, no morphing, no unwanted camera drift,
no extra people entering the frame.

That last line was added because a previous run introduced a passerby. Keeping the reason in your notes — not in the prompt — is what keeps the prompt clean.

Prefer positives for things you want

Negatives suppress; they do not create. "No blurry background" will not produce a sharp background. If you want something present, say it in the first-frame or camera slot as a positive instruction, and reserve the constraints slot for suppressing things that should not appear.

Keep it under five items per real failure category

If your constraint line starts reading like a legal disclaimer, split the problem instead: fix the camera drift with "no rotation, no shake" in the camera slot, and let the constraints tail stay short. A prompt that is mostly prohibitions has very little budget left for the motion instructions that actually make the clip work.

Iterating Recipes Without Losing the Plot

The recipes in this guide are starting points, not finished prompts. The workflow that turns them into something reliable is boring and effective: change one variable per run, and write down what happened.

Change one slot at a time

If a clip is wrong, resist the urge to rewrite the whole prompt. Pick the slot most likely responsible and change only that:

  • Motion in the wrong direction → camera slot.
  • Motion too fast or resolves too early → pacing slot.
  • Subject identity drifting → first-frame anchors.
  • Random sounds or voices → audio slot.
  • Repeated artifact → constraints tail.

One change per run means one attributable difference. Rewrite everything at once and you learn nothing, even if the next clip is better.

Keep a run log

A short table-free log per shot is enough:

shot: product_hero_01
recipe: v3
change: camera "slow push in" -> "static, no movement"
result: motion clean; subject action now reads clearly
next: try "very slow push in" for comparison

Three fields — what changed, what happened, what next — are enough to make a prompt reproducible across a team. When someone asks why a shot looks the way it does, the log answers it.

Freeze winning recipes

When a recipe stabilizes, stop editing it. Copy it into a shared library with a short name and the run log attached. New shots should start from the closest frozen recipe, not from a blank prompt. This is how a prompt set becomes an asset rather than a pile of one-off strings.

Know when to stop

Iteration has diminishing returns. If three consecutive single-variable changes produce no meaningful improvement, the problem is usually the input image, not the prompt — a first frame with an ambiguous pose or crowded background will keep fighting you. Change the still, then re-run the locked recipe.

The loop is simple: run, compare, change one thing, log it. Do that consistently and your recipes become faster to apply with every shot you finish. If you want a starting point, start with the wan2.6-flash image-to-video model page and fill in the five slots from the skeleton above.

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