AI Image-to-Video Prompts
A camera-grammar reference for any still image you want to animate on Martini. Each recipe is a motion-only prompt — push-in, dolly, orbit, parallax, drift, time-lapse — paired with the model that handles that motion best (Seedance 2 for general fidelity, Kling 3 for cinematic camera moves, Veo for natural-light photoreal, Luma Ray for ambient parallax). Use camera vocabulary, not "make it move," and chain shorter clips with last-frame extraction when you need duration past 8 seconds.
When to use this prompt
- Turning a portfolio still into a 5-second loop for IG Reels or a portfolio site.
- Animating a campaign hero on the homepage so visitors get motion above the fold.
- Motion-testing a storyboard frame to validate the cut feel before locking the shot.
- Giving a photo-only product page a motion variant for paid retargeting.
- Building an animatic from existing photography without commissioning a fresh shoot.
Required inputs
- A high-resolution still image (the model amplifies low-res artifacts across motion).
- A motion intent in plain language ("slow push-in", "60-degree orbit", "ambient parallax").
- Aspect ratio target — most models handle 16:9, 9:16, and 1:1; pick before generating.
- Duration target — most clips work cleanly at 4–8 seconds; longer needs chaining.
- Optional: a last-frame target image when you plan to chain clips for longer sequences.
Prompt recipes
Slow push-in (subtle drift)
The default cinematic motion. Subtle, repeatable, holds identity. Works on portraits, products, and environments.
Slow camera push-in toward subject in referenced image, 5 seconds, gentle parallax in foreground vs background, locked subject, subtle breathing, no rotation, no zoom-out, ends 1.2x closer than start.
Variations
- Switch to slow pull-back for reveals.
- Add subtle handheld breathe for documentary feel.
60-degree orbit
Showcases dimensionality on a single still. Best for products and sculptural subjects.
Camera orbits 60 degrees around subject in referenced image, 6 seconds, even arc speed, soft parallax background, locked focal length, no zoom, no cut, ends three-quarter view.
Tighten to 30 degrees for portraits to avoid identity drift.
Pan-and-zoom landscape
Cinematic landscape coverage on a single wide still. Veo handles the natural light shift cleanly.
Slow horizontal pan left to right across referenced wide landscape, 7 seconds, gradual zoom-in toward focal point, soft natural light shifting subtly, no cuts, ends framed on focal subject.
Variations
- Reverse direction for outro cuts.
- Replace zoom-in with zoom-out for opening establishing shots.
Rim-light reveal on hero object
Premium reveal cut for hero objects, fragrance, spirits, electronics.
Camera dollies in toward hero subject in referenced image over 5 seconds, dramatic backlight transitions to fully lit, atmosphere haze, locked subject, no rotation, ends in fully lit hero frame.
Handheld breathe
Adds presence to a static still without changing composition. Pairs naturally with voiceover.
Subtle handheld camera shake on subject in referenced image, 5 seconds, slight breathe in vertical axis, locked subject and frame, no zoom, no rotation, documentary realism.
Avoid on packaging shots — breathing causes label drift.
Dolly-in matched-on-action
Clip designed to chain into a next shot via last-frame extraction. The final frame becomes the seed for the following clip.
Camera dollies forward toward subject in referenced image, 4 seconds, motion settles by final frame to match a specific composition for chaining, locked subject, no rotation, ends framed for cut.
After generating, extract the last frame on the canvas and feed it into the next image-to-video node.
Anchored subject + environment parallax
Identity-safe motion for portraits and characters. The face stays anchored while the world breathes.
Subject in referenced image stays locked center frame, environment moves with subtle parallax — wind through background trees, distant clouds drifting, foreground bokeh shifts — 6 seconds, no camera motion on subject, no cuts.
Variations
- Swap wind for snow drift for winter scenes.
- Add light rain for moody product or portrait cuts.
Camera swap (front to three-quarter)
Director-led re-framing on a single still. Use Runway Gen-4 when you need precise motion control.
Camera transitions smoothly from front view to three-quarter view of subject in referenced image, 5 seconds, gentle arc, soft light shift through the move, locked subject, no zoom, ends three-quarter framed.
Time-lapse drift
Ambient drift cut for landscape and environmental b-roll. Luma Ray excels at slow ambient motion.
Time-lapse motion across referenced scene, 6 seconds, clouds racing, water flowing, light shifting from morning to noon to dusk, locked camera, no cuts, ends saturated golden hour.
Best at 16:9 or 21:9 cinematic crops.
Fast variant sweep (motion test)
Light prompt for testing motion intensity and pacing across multiple generations before committing render budget.
Slow push-in toward subject in referenced image, 4 seconds, soft natural light, locked subject, no rotation, no cut. Generate 6 variants for selection.
Pick the strongest variant, scale up on Seedance 2 or Kling 3.
Martini canvas workflow
Drop the still image into an image node on the canvas. If the image is below 1080p, run it through the upscale tool first — image-to-video amplifies low-res artifacts and noise across every frame.
Wire the image into the right video model based on motion intent: Seedance 2 for general subjects with reference fidelity, Kling 3 for cinematic camera moves, Veo for natural-light photoreal, Luma Ray for ambient parallax, Runway Gen-4 when motion intent is precise. Mixing models per shot intent on one canvas beats forcing a single model.
Write camera-grammar prompts only — direction, speed, arc, light shift. Avoid asking the model to introduce new objects mid-clip; that is the recipe for drift. If you need new subjects, generate a new still and chain.
For sequences longer than 8 seconds, extract the last frame of clip 1 with the frame-extraction tool and feed it as the seed image for clip 2. The hand-off is seamless because clip 2 inherits clip 1's final composition.
Run motion-test variants on Hailuo (cheap, fast), pick the winner, then scale up on the matching higher-fidelity model. Save the canvas as a template — the next image you animate reuses the entire chain.
Variations
Subtle (4 seconds)
Minimal motion — slow push-in, ambient parallax, light drift. Default for paid-social and homepage hero loops.
Cinematic (8 seconds)
Dramatic camera moves — orbit, dolly, rim-light reveal. For top-of-funnel hero placements and pre-roll.
Time-lapse (6+ seconds)
Ambient drift — clouds, water, light shift, foliage. Establishing b-roll and atmospheric inserts.
9:16 vertical
Vertical framing for IG Reels, TikTok, YouTube Shorts. Hook beat in first 1.5 seconds.
16:9 widescreen
Editorial framing for YouTube, web hero, pre-roll, TV. Room for environmental cues.
1:1 square loop
Square framing optimized for seamless looping — first frame and last frame match.
Related features
Related how-tos
Related models
Related blog posts
Related docs
Open-source prompt library
Frequently asked questions
- Why does "make it move" produce bad results?
- Image-to-video models follow camera grammar, not generic verbs. "Slow push-in", "60-degree orbit", "subtle parallax" all map to specific motion primitives the model understands. "Make it move" is ambiguous and gets averaged into chaotic motion. Use the camera vocabulary in these recipes.
- How long can a single clip run?
- Most image-to-video models hold composition through 4–8 seconds cleanly. Past 8 seconds, identity drift and motion incoherence creep in. For longer sequences, generate shorter clips and chain them with last-frame extraction so each clip seeds the next.
- Why does the subject change mid-clip?
- Two common causes: a first-frame-jump prompt (the model re-imagines the input on the first frame instead of starting from it), or a prompt that introduces new objects ("a bird flies in"). Lock the subject explicitly with phrases like "locked subject, no rotation" and only describe motion you want applied to existing pixels.
- Should I upscale the still before animating?
- Yes if it is below 1080p. Image-to-video amplifies low-res artifacts and noise across every generated frame, so a soft input becomes a soft moving output. Run the still through the upscale tool first, then wire the upscaled version into the video node.
- How do I chain multiple clips into a long sequence?
- Use the frame-extraction tool on the canvas to pull the last frame of clip 1, then feed that frame as the seed image for clip 2 with a continuity prompt. Repeat to build sequences of any length. For multi-shot continuity with a recurring character, see /prompts/video/multi-shot-video-prompts.
- Which model handles which motion best?
- Seedance 2 for general subjects and reference fidelity. Kling 3 for cinematic camera moves (orbit, dolly, dramatic reveal). Veo for natural-light photoreal scenes. Luma Ray for ambient parallax and slow drift. Runway Gen-4 when motion intent is precise. Hailuo for cheap variant sweeps before scaling up.
Try this prompt on Martini
Copy a recipe above, drop it into a node, and run it inside a full canvas workflow.