Here’s a mistake I see constantly with image-to-video. And honestly? I get it. You upload a frame you love, then carefully re-describe the whole image in the prompt — the person, the outfit, the room, the light — and tack “and then she moves” on at the end.
The generation comes back wrong in a weirdly specific way. The face has shifted. The wardrobe got redesigned. The room redecorated itself. Your prompt fought the image, and the image lost.
Look, image-to-video prompting is a different job than text-to-video prompting. Treating it like the same job is where the waste starts. In my pipeline, image-to-video is the default, because my workflow is storyboard-first: frames get approved as stills before any motion is generated. So by the time I’m writing an I2V prompt, the frame has already been argued over, fixed, and locked.
The prompt has exactly one job left. Choreography. That’s what this guide is about — writing prompts that do that one job well.
The Image Already Said It
An uploaded frame is the most detailed prompt you’ll ever hand a model. Seriously. Every pixel is an instruction: this face, this jacket, this wallpaper, this exact afternoon light through that exact window. So when your text describes the same things again in words, you’re not reinforcing the image. You’re offering the model a second, vaguer source of truth and asking it to reconcile the two.
And reconciliation, in a generative model, looks like drift. That’s the technical term for “it changed stuff I didn’t ask it to change.”
Here’s the mental shift that fixed my I2V work. In text-to-video, the prompt builds the world and then moves it. In image-to-video, the world is built — you’re holding it — and the prompt only moves it. Subject description, setting description, style description: the image carries all of it, at way higher fidelity than your words ever will. What the image can’t carry is time.
Motion, camera behaviour, sequence, duration, end state — that’s the entire cargo of an I2V prompt. And it’s a lighter, more powerful prompt for it.
This is also why I2V pairs so naturally with a storyboard workflow. The board settles every visual argument while settling is cheap — as stills, reviewed, compared, approved. Then generation stops being an act of imagination and becomes an act of animation. The contrast between the two modes is covered more fully in the site’s text-to-video versus image-to-video comparison, but the working consequence is simple: if you approved the frame, don’t re-describe the frame.
Animate it.
One exception, and I’ll name it honestly. Some details in a frame are fragile and worth restating in words as anchors — a logo that must not warp, a specific prop the motion will touch, text on a sign in a shot where text matters. Restate those surgically. One or two items, named exactly. That’s anchoring. Re-describing the whole image is something else entirely: it’s doubt, typed out, and the model bills you for it.
What The I2V Prompt Must Carry: Motion, Camera, Time
Strip an I2V prompt down to the three things only it can say. Write those three well.
Motion: What Moves, How, And In Which Direction
Name the moving element, the verb, the direction, and the scale of the movement. “She turns her head toward the window, slowly” — subject, verb, target, pace, nine words. That’s a complete motion instruction.
Now compare “she moves naturally with lifelike motion.” The first one tells the model what to animate. The second asks it to improvise and then grades it on naturalness it can’t even measure. See the difference?
Small motion, stated plainly, beats ambitious motion stated vaguely. Every time. A head turn, a hand lifting, a step forward, steam rising, a curtain shifting — these complete inside a short generation and hold their end states. A whole routine doesn’t.
If the storyboard needs a routine, it needs more panels. Which means more frames, which in I2V means more generations, each with its own still. The frame count is the choreography budget. No way around it.
Camera: Still, Or Moving On Purpose
The camera decision belongs in the I2V prompt because the image can’t make it. A still frame is compatible with a locked-off shot, a push-in, a pan — the image only shows where the camera starts. So say it. Explicitly. Even when the answer is “don’t move.”
“Locked-off camera” is an instruction I type constantly, and it’s never wasted. Unspecified cameras drift. And drift is motion you didn’t choreograph competing with motion you did. That’s a fight you don’t want.
When the camera does move: one move, named, with its speed. “Slow push-in” is a complete thought. “Dynamic cinematic camera movement” is an invitation for the model to audition several moves inside one clip — which it will, trust me — and none of them will match the next shot in your sequence. The detailed vocabulary for this lives in the camera movement prompting guide; the I2V-specific point is that the camera line may be half your entire prompt.
And that’s fine. Proportion follows importance, not habit.
Time: How The Motion Unfolds Across The Clip
Time language is the third cargo item, and honestly it’s the most neglected. When does the motion start — immediately, or after a beat of stillness? Does it ease in or start at full speed? Does it settle at the end or finish mid-gesture?
Phrases like “after a moment, she turns,” “movement settles by the end,” and “holds still for the final seconds” give the model a timeline instead of a loop. And they give you, in the edit, a clean frame to cut on.
That last part matters in production more than in any single generation. A clip that ends in a held state can be cut anywhere near its end. A clip that’s still mid-motion at the final frame? That forces the cut to a precise frame, or forces an ugly freeze. Neither is fun.
When I board a sequence, the end state of each panel is drawn — which means the I2V prompt inherits an obligation: arrive here, then rest. Generations that rest on cue are the ones that make a Premiere timeline feel assembled rather than glued together.
What To Leave Out (And Why It Hurts)
Knowing what to omit is most of I2V craft, honestly. So let me walk through the usual suspects, and what each one costs you.
Full subject re-description. Covered above: it creates a second, blurrier source of truth. The visible symptom is identity drift — the same person, five percent redesigned, in the exact shot where you supplied their photograph. The fix is total: if it’s visible in the frame, don’t type it.
Style and quality bait. “Cinematic, 8K, ultra-detailed, masterpiece” in an I2V prompt asks the model to re-render the image toward a style target. Which means re-interpreting the very pixels you approved. If the frame’s look is wrong, fix the frame — regenerate the still, adjust it, grade it — before motion. Motion generation is the most expensive place to have a style argument, and you’ll have it frame by frame, at video prices. Nobody wants that.
Setting description. The room is in the picture. Describing it again risks the model rebuilding it mid-clip — which is how backgrounds breathe, furniture migrates, and doorways relocate between second two and second six. If the shot needs a background change, like a door opening to somewhere new, that’s a motion event and should be written as one: “the door behind her swings open to a dark hallway.” Event, yes. Inventory, no.
Emotion retellings. Her face is in the frame. The expression is already cast. An I2V prompt that adds “she looks happy and carefree” asks the model to re-perform a face it was handed — and re-performance is where expressions inflate into melodrama. If the frame’s expression is right, write the motion and leave the face alone. If the expression must change, that change is the motion, and it should be written physically and small: “her half-smile fades.”
Lighting re-description. Same principle, slightly different symptom: relighting artefacts. The image’s light is coherent because it was rendered once, whole. Asking the model to also honour “warm golden-hour light from the left” invites it to adjust the light as the subject moves, and moving subjects under re-negotiated light flicker.
Where light genuinely matters to the motion — a face turning into shadow — write that as an event: “she turns, her face passing into shadow.” The site’s lighting prompts guide covers text-to-video lighting in depth; in I2V, light is inherited, not requested. That’s the rule.
The honest summary: every sentence in an I2V prompt that isn’t about time is a risk. The prompt should feel almost embarrassingly short next to your text-to-video prompts. That shortness isn’t laziness. It’s respect for the frame you already approved.
Describing Endpoints: Where Motion Should Land
If motion is the prompt’s job, endpoints are its signature. A movement without a stated destination is a suggestion. And models complete suggestions in the direction of maximum drama — toward camera, into a spin, out of frame. A movement with a stated destination is choreography. Big difference.
Endpoints come in three useful flavours. Spatial endpoints name where the subject finishes: “steps forward to the counter,” “turns to face the window,” “sits down on the edge of the bed.” Pose endpoints name the final arrangement: “ends with both hands resting on the table,” “head tilted, listening.” Rest endpoints name the settle: “movement eases to a stop,” “holds the final pose, breathing.”
A strong I2V prompt usually carries one spatial or pose endpoint plus one rest instruction. That combination gives the model a target and gives you a cuttable final frame. Both matter.
Endpoints are also the fix for the most common I2V disappointment: the clip that starts beautifully and then keeps going, inventing a second, unwanted beat because the first beat finished early and the duration had to be filled. Models fill time. If you don’t assign the end of the clip a job — holding, settling, breathing, gazing — the model assigns it one. And its taste in second beats is worse than yours, I’ll be straight with you.
“Then holds, looking out the window” is three words of insurance I buy on almost every generation. Cheap insurance. Worth it.
There’s a storyboard consequence here that’s easy to miss. If your workflow boards panels, the next panel is the endpoint of the current one, drawn. The generation in between is just the journey from board A to board B. Making that relationship explicit in the prompt — “ends facing the door, hand raised to knock” because the next panel shows the knock — is what turns a set of clips into a sequence. The frames stop being cousins and start being consecutive.
Negative Space: Where Movement Is Allowed To Go
Every frame contains regions where motion is welcome and regions where it’s dangerous. A good I2V prompt manages that geography, even if it never uses the words “negative space.”
Motion should be invited where the frame has room and forgiveness: open sky, water, foliage, steam, out-of-focus backgrounds, loose clothing, hair. These areas absorb generative wobble gracefully. Motion should be fenced off where the frame has precision: faces in repose, hands at rest, text, logos, product surfaces, straight architectural lines, reflections.
In those regions, stillness is the goal. And the way to get it is to give the prompt no verb that applies there — and give the moving verb a clearly different home.
Practically, this means assigning motion by name. “Steam rises from the cup; she remains still, hands around the cup” divides the frame into a moving zone and a held zone. The model, told exactly where life is allowed, largely respects the border. The failure version is “a cosy café scene with lively atmosphere” — which licenses motion everywhere and collects wobble from the whole frame, including her face. Yeah. Don’t do that.
Negative space also means literal space: leave room in the frame for the movement you’re requesting. A subject framed tightly at the right edge cannot “turn and walk left” without exiting or colliding with the border. This is a boarding problem before it’s a prompting problem, and it’s worth checking at the still stage: does this frame contain the journey?
If not, the fix is a wider still, not a cleverer prompt. I’ve re-rendered more frames for headroom than for any other single reason, and it’s saved more generations than it cost. Hands down.
Finally, allow the background its own cheap life. A gentle, explicitly background motion — “distant traffic drifts past the window, soft and out of focus” — can sell a held foreground performance, because the contrast between moving world and still subject reads as intention. The trick is depth language: distant, soft, behind. Motion assigned to the background plane stays there more reliably than motion assigned to “the scene” as a whole.
Five Motion-Only Prompt Patterns
These five patterns cover most of the I2V shots in my own projects. Each assumes the frame is approved and says nothing about how it looks. Adapt the verb and the endpoint; keep the structure.
1. The Single Gesture. “She turns her head toward the window and holds her gaze there. Locked-off camera. Movement settles, still at the end.” One gesture, one target, one rest. The workhorse for reaction shots and story beats.
2. The Short Journey. “He walks forward two steps to the counter and stops, hands resting on its edge. Slow tracking camera keeping pace.” Distance is stated in steps, not metres, because steps are body-scale and models judge bodies better than measurements. The stop and the hands give it a drawn endpoint.
3. The Ambient Frame. “Rain falls steadily outside the window behind her; she stays still, breathing slowly, cup held in both hands. Static camera.” All motion is exiled to texture, the subject is pure held state. Useful for pauses, listening shots, and any frame whose face you can’t afford to have re-performed.
4. The Camera-Only Move. “Slow push-in from medium shot to close-up. Subject holds position, only breathing and a slight blink.” When the frame is perfect and fragile — a product hero, a face you finally nailed — move the camera and nothing else. This pattern has rescued more approved stills in my pipeline than any other. Seriously.
5. The Object Beat. “The envelope slides across the counter toward camera and comes to rest. Locked-off camera, focus on the envelope.” Objects don’t have faces to drift, and a sliding, pouring, opening, or settling object makes a clean cutaway that covers edits elsewhere. Cutaways generated this way are cheap, reliable, and cut like live action.
One closing habit that ties the five together: before generating, read your I2V prompt and delete every word the image already shows. What remains should be a short paragraph about time — what moves, where the camera goes, where it all lands, and what rests.
If nothing remains worth saying, the frame wants to be a still in the edit. And there’s no shame in that. Some of the strongest beats in a sequence are approved frames held for two seconds over sound.
Not every panel has to move. The ones that do will move better for having a prompt that only ever talked about motion.