You have an image you love — a character portrait with perfect lighting, a landscape that took twelve attempts to get right — and it just sits there. Regenerating the whole scene as a video means rolling the dice again on everything you already nailed. Image-to-video solves exactly this: it takes your finished image as the first frame and animates forward from it.
This guide covers when image-to-video beats text-to-video, how to choose a source image that survives being animated, and how to write the motion prompt. Everything here works in Imageny's video generator.
What is image-to-video AI?
A text-to-video model invents everything from your prompt: subject, style, lighting, and motion. An image-to-video model is handed frame one and only has to invent what happens next. Your composition, character design, and color grading are locked in before any motion is generated.
That trade shapes what each mode is good at:
- Use text-to-video when the motion is the point and you don't care about an exact starting look — a drone shot over mountains, an abstract loop. Our text-to-video guide covers that workflow.
- Use image-to-video when the look is the point: a specific character, a scene you already art-directed, or a style you spent real time dialing in.
The practical difference shows up in iteration cost. With text-to-video, a bad result means re-prompting the entire scene. With image-to-video, a bad result only means re-prompting the motion — the frame you approved never changes.
What makes a good source image?
Not every image animates well. Three properties predict success:
A clear subject with room to move. A tight close-up crops off the shoulders and hair the model needs to animate a head turn. Medium shots and full-body compositions give motion somewhere to go. If you're generating the source image fresh in the playground, frame it a step wider than you would for a standalone image.
Coherent anatomy and structure. Whatever is slightly wrong in the still — a suspicious hand, a tangled hair strand — gets worse once it moves. Warping compounds frame over frame. Fix flaws in the image stage, where a re-roll is cheap, not the video stage, where it costs a full generation.
Simple backgrounds. Crowds, dense foliage, and busy signage give the model hundreds of small elements to keep consistent, and small elements are where flicker starts. A character against sky, an interior with a few large shapes, or a soft bokeh background all hold together far better.
How do I write a motion prompt for an image?
Do not re-describe the image. The model can already see the silver hair, the neon alley, the rain. Re-describing the scene at best wastes tokens and at worst nudges the model to "correct" details you wanted kept.
Describe only what changes. A reliable structure is one subject motion plus one camera move:
she slowly turns her head toward the camera and smiles, gentle push-incherry blossom petals drift across the frame, camera holds stillsteam rises from the cup, slow orbit around the tablehis coat moves in the wind, handheld shot, subtle sway
Keep both motions small. Image-to-video is at its best with restrained, atmospheric movement — hair in wind, drifting particles, breathing, a slow camera drift. Ask for a backflip and the model has to invent every intermediate pose from a single reference frame, which is where limbs go wrong.
Why does my clip warp or melt, and how do I fix it?
Almost all image-to-video failures trace back to three causes, each with a direct fix:
- Too much motion requested. "She runs through the crowd dodging traffic" collapses fast. Cut it to one motion: "she walks forward, crowd blurred behind her."
- The subject leaves the frame. Once a character walks out of view, the model must reinvent them if they return, and the reinvention rarely matches. Keep motion inside the frame, or let the camera follow the subject.
- A flawed source image. If frame one has a broken hand, frame forty has a broken arm. Go back and regenerate the still first.
When a clip is close but not right, change only the motion phrase and run it again. Because the source frame is fixed, iterations are controlled experiments — you're testing one variable, not re-rolling the whole scene.
A workflow that works end to end
- Generate the still in the playground — subject, style, lighting, composition. Iterate here until it's genuinely right; this is the cheap stage.
- Frame a step wider than feels natural, so motion has room.
- Move to video generation with the image as your source.
- Prompt one subject motion and one camera move, both modest.
- Iterate on the motion phrase only.
The habit that changes results the most: treat the still image as the deliverable and the video as a pass on top of it. Creators who polish the frame first spend their video credits refining motion, not repairing anatomy.
Try it on your own image
Pick the best image in your gallery — or make one in the playground — and animate it in Imageny's video generator. Start with one small motion and one slow camera move, and you'll have a clip worth keeping on the first or second attempt.
Cover photo by Avel Chuklanov on Unsplash.