Prepare a Source Image for AI Video: Crop, Detail and Motion

Start with a shot requirement, not a resolution target

Write a one-sentence purpose for the clip. For example:

A short vertical product shot should draw attention to the bottle while keeping its complete outline and capacity marking visible.

That sentence is more useful than “make a cinematic video.” It identifies a subject, destination and acceptance condition. Next, choose the smallest visible event that serves it: a gentle background change, a restrained camera approach, or an action by the subject.

These choices need different source material. A background-light effect can use a stationary front view. A person raising a hand needs enough of the arm and surrounding space. A camera moving around a building asks for surfaces and depth relationships absent from one photograph.

The motion-choice guide compares subject, camera and atmospheric movement under one brief. For preparation, your job is to establish what the source can support before choosing a more demanding movement.

Create three short lists:

“Recognizable” and “fully visible” are separate requirements. A bottle may still look like the same bottle after its capacity text has left the frame. Decide which standard your project needs.

Inspect the details the viewer will rely on

Open the best available original at a useful size. Inspect the subject closely, then view it at the approximate size at which people will watch the video. The first check reveals defects; the second tells you which details communicate the intended message.

For a portrait, examine eye shape, hairline, glasses, mouth, visible teeth and the outline where the face meets the background. For a product, inspect its silhouette, label, opening, handle or other defining feature. For scenery, follow the horizon, rooflines, window rows and foreground overlaps.

Runway's current image-to-video guidance warns that source artifacts, including blurry faces or hands, can become more pronounced in video. It also distinguishes the image's visual information from the prompt's motion instructions. That is useful preparation advice; it does not establish a result for your particular Dreamina image. Runway image-to-video prompting guide.

Classify what you find before editing:

• Distracting but nonessential: a small background blemish can be repaired or cropped if doing so preserves the intended composition.
• Essential and inaccurate: a wrong date, misspelled label or distorted eye should be corrected and approved before animation.
• Essential but unavailable: a hidden hand, unreadable original label or unseen side of a building calls for another source or a less revealing shot.

Sharpening a blurred word does not establish what the word originally said. Generating a plausible hand does not recover the person's actual hand position. If the missing information matters, obtain a better source rather than treating a convincing reconstruction as verification.

Keep the original beside the prepared version. After any image edit, compare the features that made the image useful. A cleaner background is not an improvement if the same edit changes the face or product.

Give movement space in the destination frame

Prepare for the final shape before spending credits. A photograph that works horizontally may lose its subject when cropped vertically. Check the intended composition as an actual rectangle, not an imagined future adjustment.

Consider a hypothetical 4000 × 3000 landscape image prepared for a 9:16 clip. A full-height 9:16 crop is about 1688 pixels wide. It retains roughly 42% of the original width. If two people stand far apart, both may not fit; if a building fills the width, its side wings may disappear.

Your options are practical:

1
1. Move the crop if all essential content still fits.
2
2. Choose another photograph with a more suitable composition.
3
3. Use a wider output if the destination permits it.
4
4. Build a deliberate layout around the intact image instead of filling the screen with a destructive crop.
5
5. Extend the surroundings only when newly invented context is acceptable, then inspect the extension before animation.

Do not stretch a subject to make it fit. Extending the canvas also does not create verified scenery outside the photograph.

Check the end of a push-in before committing to its beginning

For a conventional, centered zoom, increasing scale reduces the part of the image visible inside the frame. At 110% of the starting scale, about 90.9% of the starting width and height remains visible. Roughly 4.5% disappears from each side when the crop is centered.

This calculation is a preparation aid for a flat-image zoom. A generated camera move can change framing in additional ways; a prompt saying “slow” does not define an exact 110% scale.

Temporarily preview a tighter crop before generation. Does the bottom label survive? Does a hand leave the frame? Is the person's forehead already at the upper edge? If so, widen the source, reduce the intended movement or choose a fixed-camera action.

There is no universal safe margin. The useful margin depends on the subject's starting position and the greatest movement you allow. Check the least spacious edge, not only the average space around the subject.

What a recorded bottle animation reveals about source preparation

The following example uses the saved prompt, settings and original output frames from a recorded browser demonstration.

The source showed a complete green bottle, including its open neck, three labels, red star and “500 ml” marking. A five-second Dreamina Seedance 2.5 generation requested a very slow push toward the stationary bottle while keeping those features readable. The source and returned clip used a 3:4 composition.

The opening output frame includes the whole bottle. At 3.75 seconds, the capacity marking is outside the frame; by the five-second checkpoint, much of the red star is also cropped out. The motion happened, but full packaging visibility did not survive it.

Opening and five-second frames from one recorded Dreamina animation, October 14, 2026. Source photograph: Philafrenzy, “San Pellegrino 500ml bottle,” Wikimedia Commons, CC BY-SA 4.0. These AI-modified frame adaptations are shared under the same license. The branded bottle is an illustrative test subject.

The saved review covered all 121 video frames as thumbnails and five checkpoints at the original 834 × 1112 output size. That supports the observed framing problem; it is not full-resolution certification of every letter throughout the clip.

The preparation lesson is specific: a readable source is insufficient when the motion removes the readable area. A wider composition or smaller move is a proposed correction, not a successful retry demonstrated here. If every label detail must remain fixed, decide that before using generative motion and consider a supplied product layer in the final edit.

Check what the movement would have to invent

A single image establishes one visible arrangement. Use it to identify the boundary between information supplied by the source and information the shot would require later.

These are planning implications, not predictions of a particular model's failure rate. A generator may produce a plausible new view. Plausible does not mean it matches the real unseen object.

Also inspect occlusion: a hand over a face, a scarf crossing a mouth, a branch hiding a roofline. When one object moves away, something must fill the space it previously covered. A source with fewer critical overlaps is often easier to approve because less important information must be reconstructed.

For buildings and landscapes, the geometry-preservation guide develops this into a region-by-region motion plan. Use that approach before requesting a lateral move through a scene with strong foreground depth.

Look for motion already implied by the still

The source may suggest an action even before you write a prompt. Windblown hair, a lifted foot, splashing water or a blurred wheel can make a stationary brief ambiguous.

Runway documents this issue using implied motion cues such as blur, mid-action poses and directional elements. Its recommendation is to inspect conflicting cues when the requested movement is difficult to obtain. This is a model-specific example of why source and motion should agree, rather than evidence that every such image will fail. Runway prompting guidance.

If you want a quiet portrait, select a settled pose rather than an expressive mid-gesture photograph. If you want a parked car, prefer a clean stationary view over a racing image full of spray. If the image is the only one available, let the brief acknowledge its pose instead of adding contradictory instructions.

Avoid cleaning away meaningful identity cues merely to simplify the picture. A distinctive hairstyle or accessory may be part of what the viewer recognizes. Change the movement first when source repair would change the subject you are trying to preserve.

Prepare one approved image and an explicit input role

Once you choose the source, finish the still before generating motion:

1
1. Preserve an untouched original.
2
2. Correct factual mistakes and distracting defects in a working copy.
3
3. Apply the intended crop and review the tightest permitted framing.
4
4. Compare face, product or scene anchors with the original.
5
5. Export one approved working image and give it an identifiable filename.
6
6. Record the motion, fixed details and destination alongside that file.

Use Dreamina's image tools when the source needs a creative revision, but approve their output as a new still. Do not rely on the video stage to decide whether a changed label or reconstructed face is acceptable.

The input's role matters too. An image uploaded as a general reference can guide subject and appearance; a first-frame input is intended to establish the opening composition. They are not interchangeable promises. Dreamina's current Seedance 2.5 guide documents AI Video, Seedance 2.5, Omni reference and asset tags. Follow the controls available for the selected route, and explicitly identify what your image supplies. Dreamina Seedance 2.5 guide.

Before submitting, confirm that the actual preview contains the approved crop. Check accepted file requirements in the current upload control and read the generation cost for your chosen settings. There is no single pixel threshold or file-size rule that this guide can apply to every model and account.

Use a preparation card to make the first attempt easy to judge

Here is an illustrative card for a mug photograph. It is a proposed planning example, not a tested generation.

An example instruction for a reference-based route would be:

Use @Image1 for the mug, table and daylight setting. A narrow curl of steam rises slowly above the rim. The camera remains fixed. Keep the mug outline, handle position and painted pattern consistent with the reference. One continuous shot.

Replace the asset tag with the one shown in your workspace. If using a dedicated first-frame route, supply the approved image in that slot rather than assuming the tag makes it a first frame.

The card keeps the result assessable. “More beautiful” is difficult to diagnose; “the handle changes halfway through” points to a specific revision. It also prevents a later reviewer from judging a mood study as though it were a factual product demonstration.

Repair, replace or change the shot?

Use the first visible obstacle to choose the next action.

A conventional pan or zoom is suitable when the photograph itself only needs to move within the frame. It does not animate a person's expression or make water ripple. Adobe documents still-image preparation and conventional Position/Scale controls; prepare sufficient pixels for the largest planned view rather than expecting a zoom to create detail. Adobe still-image preparation.

Finish preparation when you can hand someone one approved image, its intended crop, one observable action and a short list of details they must check. Then generate a short draft and use the full-clip review workflow to decide whether the actual movement meets those requirements. A strong source makes that decision clearer; the saved video remains the final evidence.