Skip to content

Tools

AI Image to Video

See how a starting image becomes a five-second clip. Then upload your own photo and describe what should move.

Upload a photo to animate
Starting image of a cat courier on a hovering scooter

Image to video · Cat courier

Starting image and result · Seedance 1.0 Pro Fast · 5s

This clip used 41 video credits. Check the estimate for your run before generating.

Choosing the source still for AI image to video

Most of the quality of an image-to-video clip is decided before the video model runs, by the still you attach. Four properties matter. Subject separation: the thing that should move must be unambiguous, which usually means one dominant subject and a simple background, because a busy frame gives the engine too many things to reinterpret. Edge quality: crisp silhouettes survive interpolation better than soft, busy ones.

Lighting direction: a still with clear directional light gives the model something consistent to move, while flat ambient light invites it to invent shadows. And headroom: leave a little empty space in the direction of the camera move, because a clip that pushes in needs somewhere to push into. If you are choosing between two approved stills, pick the one with the simpler silhouette, not the prettier one.

The motion budget: what one second can carry

Video models do not render motion; they render plausible change between frames, which means there is a budget and exceeding it produces drift. A safe rule is one clear movement per second, and for a first clip, considerably less than that. Count the things you asked to move: the subject, the camera, the fabric, the hair, the background, the lighting. Every extra moving element divides the same budget. A subject turn plus a camera push plus drifting fabric is already three.

Ask for two and you will usually get both; ask for five and you will get none of them cleanly. This is also why longer clips drift more than short ones at the same prompt: a five-second clip makes fifteen independent decisions where a two-second clip makes six. Keep clips short and specific, and stitch them.

Why clips drift, and how to re-anchor

Drift has four common shapes, and each has a different fix. Identity drift: the subject's face or product shape changes after the first second. Fix it by shortening the clip and by using an engine that treats the attached still as a hard first frame rather than loose reference. Geometric drift: straight lines bow, tiles warp, text on a label smears. Fix it by removing the straight-line content from the shot, since video models cannot hold parallel structure over time.

Lighting drift: shadows move or invert because the model re-lights each frame. Fix it by naming a fixed light source in the prompt. Physics drift: cloth, water, and hair ignore gravity and inertia. Fix it by removing them. When a clip drifts, the cheapest repair is usually to re-anchor — regenerate with the same still as the first frame and one added constraint — rather than to keep re-rolling the same prompt.

Chaining stills into a sequence

A sequence is built from separate shots, not from one long generation. Generate each still separately at the final ratio, with the palette, lens and light clause identical across shots, then animate them one at a time and cut them together. Keep each shot inside its own motion budget — two seconds for a move, four for a settle — and let the edit, not the model, create continuity.

The payoff is reliability: a sequence of four clean two-second shots is far more usable than one eight-second take that drifts in the middle. The cost is that you have to shoot each still, which is exactly the same budgeting discipline you already use for images. If a sequence needs a long take with real continuity, that is what text-to-video is for, and the two approaches should be chosen per shot rather than per project.

Matching the first and last frame

If you want two clips to cut cleanly together, generate them so that one ends where the next begins. The practical method is to make the still for shot two from the last frame of shot one, or to prompt shot one with an explicit final position — the subject facing left, hand on the table — and prompt shot two with that same position as its opening.

Engines that expose a last-frame or end-frame control do this for you, and when one is available, use it rather than trying to describe the position in words. When no control exists, the fallback is a slower clip that ends near a rest position, since a moving frame is very hard to match and a settled one is easy. Then check the two frames side by side before you spend anything on the second clip.

Camera-move vocabulary that survives

Named camera moves work better than descriptions of them, because the engine has been trained on the terms. These survive reliably: slow push in, pull back, static shot, pan left or right, tilt up or down, orbit around the subject, and handheld follow. These survive less reliably: crane up, whip pan, dutch angle, zoom with a moving subject, and anything combining two moves. Two habits make the difference.

Describe the move in one clause, and add speed and direction — "slow push in from the chest up" rather than "cinematic camera movement". And state what must not move, because the model's default assumption is that everything may: "subject stays seated, background unchanged, no camera shake". The negative half of the sentence is often worth more than the positive half.

When to stop at stills and animate a photo instead

Image-to-video is the right tool when the camera moves and the subject stays recognisable, and the wrong tool for several common jobs. If the goal is a portrait that moves subtly — a blink, a breath, a slight turn of the head — you want a photo animation workflow rather than image-to-video, because those engines are tuned to preserve a face rather than to move a camera through a scene. If the goal is to bring a still photograph to life, that is again the photo animation lane.

If the goal is a product turntable or a logo sting, image-to-video from a clean render is exactly right. And if the shot needs a person whose identity must stay fixed across a long take, generate the identity once, hold that still, and cut rather than generate long.

Which engine for which kind of motion

Match the engine to the kind of movement rather than to the brand. Fast, physically plausible motion — people, animals, product handling, natural camera moves — is where Seedance and WAN are strongest, and they are the sensible default. Cinematic, high-production shots with strong art direction are where Sora and Veo earn their cost, and they are worth it for a hero shot rather than for twelve social cuts.

If your source still is a real photograph, the deciding factor is often not quality but whether the engine accepts a reference at all, and several will reject an uploaded still outright — in that case switch engines rather than re-exporting the image, since the file is rarely the problem. Keep the still in your library so re-running on another engine costs you one job, not a new shoot.

What AI image to video costs by duration and resolution

Duration and resolution are the two settings that move the bill, and both interact with drift. A four-second 1080p clip costs roughly twice a two-second 720p clip, and buys you twice the opportunities for something to go wrong. Shoot drafts at short duration and standard resolution, then pay for the final only once you have a clip whose motion survives the full length. Two habits save real money.

Cap your draft duration at the shortest length that could plausibly be the final cut, and if the final needs to be longer, generate the long version as several short clips rather than one long one — which is also what you would do for reliability reasons. Image and video credits share one balance, so an afternoon of stills and an evening of clips are competing for the same budget; decide which one the day is for.

A one-clip production workflow

Minutes 0–5: write the shot as one sentence — subject, movement, camera, duration. Minutes 5–15: generate the source still and approve it, because an unapproved still is the most expensive mistake in this workflow. Minutes 15–25: draft the clip at the shortest duration on one engine, with the negative constraints written explicitly. Minutes 25–35: if it drifts, re-anchor once with one added constraint; if it still drifts, cut the motion list rather than re-rolling.

Minutes 35–45: run the winner at final duration and resolution on the second engine and compare. Save the still, the prompt, and the accepted clip together in your library, so the next shot in the sequence starts from a known-good reference instead of a fresh experiment.

Fixing an AI image to video clip instead of regenerating it

Before you spend another generation, ask whether the defect is fixable downstream. Colour casts and exposure problems are almost always a grade, not a regeneration — fix them in an editor for free. A slight identity shift at the end can be covered by trimming the last half-second, which is often the cheapest fix available in video work. Speed is a re-time, and a slow push-in that runs slightly fast can be slowed in post.

What is not fixable downstream: a warped face, melted geometry, or a subject that changes shape. Those mean re-anchoring, and re-anchoring means a new generation. Sort each failed clip into fixable or not before you re-roll it, because the fixable ones are the ones quietly eating credit budgets on failed retries.

What a shot list costs before you generate

Budget from the shot list, not from the video. A ten-shot social cut is ten stills plus ten clips, and the stills cost roughly the same as the clips at short durations — which surprises people who budget video as one line item. Add the re-rolls you should honestly expect: a first clip that works without adjustment is the exception, so plan for one and a half to two generations per shot on the first engine. If a second engine is used for comparison, that is another shot.

The practical control is shot count, not duration: six shots of two seconds with two attempts each is a defensible budget, while one ten-second hero shot with ten attempts is not. Write the shot list down first — it is the only part of the process that costs nothing and prevents the expensive kind of enthusiasm.

Open studio

FAQ

Which models work for image-to-video?
Start with Seedance or WAN when they show reference / first-frame controls in Video Studio. Switch engines if a job rejects your still.
Why did my job reject the image?
Not every video model accepts references. Switch to an engine that shows reference controls in Video Studio, or generate text-to-video instead.

Related models

Compare