How to Animate an AI-Generated Still Without Warping the Subject

The main workflow below produces a five-second camera move in a video editor. There is also a command-line recipe, guidance for separating background motion, and a generative branch for shots that need new action. You can practise with the ceramic hummingbird still and compare it with the original five-second example.

Choose what should move before choosing a tool

A push-in enlarges the existing picture. A true camera orbit reveals surfaces that the still does not show. Those are different jobs, and the second requires additional views, a 3D asset, or invented image content. Decide which one your shot needs before spending time on prompts.

For a product label or a person's likeness that must stay accurate, begin with the complete-image method. Add more complex motion only when it serves the shot. Keeping width and height linked preserves proportions. Resizing and video compression can still soften small details or change color slightly; preserving shape does not mean preserving every pixel value.

Prepare the still and mark the details that must survive

Inspect the source at full size before importing it. Check hands, eyes, lettering, repeated patterns, and the outline of the subject. Correct source defects before animation, then save that accepted image separately. Runway's image-to-video guidance warns that visual artifacts such as blurry hands and faces can become more pronounced in video.

Write a short list of visible details to compare later. Choose details that reveal different failures: the number of buttons, the shape of a rim, a corner, and the outline of a face. "Looks like the same person" is too broad to help locate a change.

The hummingbird example has a metallic-gold beak, exactly three white dots on the near wing, a triangular chip at the front-right corner of the base, and a cobalt silhouette. The still is 1672 × 941 pixels. These features make the example easy to check; they are not instructions you need to copy into every video prompt.

For your own image, record the details in plain language and note any that sit close to the frame edge. A face can keep its shape while a zoom cuts off the chin. That is a framing failure, and it needs a different fix from a face that changes shape.

Set the output shape before animating

Choose the aspect ratio first. For this landscape exercise, use 1920 × 1080 at 30 frames per second. For a vertical deliverable, compose in 1080 × 1920 from the beginning. Reframing after animation can remove a face, a label, or the subject's base even when the original render looks correct.

Use a proportional fit or fill operation. Fit keeps the entire image and can leave borders. Fill covers the frame and crops part of an image with a different aspect ratio. Stretching width and height independently changes the subject's proportions before the animation even starts.

A 1920 × 1080 image cropped to a full-height 9:16 view retains only about 608 pixels of its original width, roughly 32%. If the subject is wider than that strip, keep the complete picture inside the vertical canvas with a designed background, or prepare a separate vertical composition. Generative expansion can add surroundings, but it creates new content that needs inspection near the subject.

Allow enough resolution for the zoom

Check the tightest crop, not only the first frame. A 5.5% push-in on a 1920 × 1080 composition needs roughly 2026 × 1140 source pixels to avoid enlargement, assuming the source already has the same aspect ratio. The downloadable 1672 × 941 example is enlarged for its 1080p render. It demonstrates stable geometry; it does not demonstrate native 1080p detail.

If the closest view looks soft, reduce the zoom, export smaller, or return to a larger accepted source. An AI upscaler can also change small marks, skin detail, and lettering. Compare its output with the original before treating the upscaled image as your new source.

Make a five-second push-in with scale and position keyframes

Use an editor that can animate an image's scale and position while keeping width and height linked. In Premiere, select the image clip, choose Window > Effect Controls, and expand Motion, as described in Adobe's Motion instructions. Enable Toggle Animation beside Scale and Position to create keyframes. A keyframe stores a value at a particular moment; the editor calculates the values between keyframes. Other editors may use different labels. The percentages below are relative to the image's initial fitted size.

1. Import the accepted still, place it on one video track, and give it a five-second duration. Turn off automatic animation presets so you can control the movement.
2. Fit or fill the image proportionally, center it, and inspect the crop. Treat this fitted size as 100%. Keep a fixed center anchor point throughout the move.
3. At the first frame, add keyframes for uniform scale and position. Use 100% relative scale and zero offset from the centered position.
4. At frame 128, about 4.27 seconds (timecode 00:00:04:08 at 30 fps), set relative scale to 105.5%, horizontal offset to −18 pixels, and vertical offset to −8 pixels. In a 1920 × 1080 coordinate system with the center at 960, 540, that final center is 942, 532.
5. Hold those values through frame 149, the last frame at about 4.97 seconds (00:00:04:29 at 30 fps). The clip occupies five seconds; there is no encoded frame at exactly 5.00 seconds in this 150-frame example.
6. Start with linear interpolation on all properties. If you add easing, use matching timing and progress curves for scale and position, then scrub the whole move. A faster pan combined with a slower zoom can expose a border midway through the shot.
7. Export a silent H.264 MP4 at 1920 × 1080 and 30 fps. Watch that file and check its first and last frames, rather than relying only on the editor preview.

If your editor displays scale relative to native image size, multiply its starting value by 1.055. A fitted value of 114.8% would end near 121.1%. Entering 105.5% directly in that case would shrink the image instead of making the intended push-in.

These numbers reproduce the design of the archived hummingbird example. For a different picture, begin with a centered zoom and add a small pan only after checking the crop. You can omit the pan entirely when the subject already fills the frame.

Check the border clearance mathematically

For a centered image that exactly covers a same-aspect-ratio frame, uniform scale creates extra image area beyond each edge. At any moment, the translation must stay within that extra area:

horizontal clearance per side = frame width × (scale − 1) / 2
vertical clearance per side   = frame height × (scale − 1) / 2

At 1920 × 1080 and scale 1.055:
horizontal clearance = 52.8 pixels
vertical clearance   = 29.7 pixels

The example's final offsets of 18 and 8 pixels fit within those limits. At 100% scale there is zero extra clearance, so panning immediately at that scale exposes an edge. Let zoom and pan advance together: halfway through the move, both should be halfway to their final values. You can also begin slightly enlarged or reduce the movement. If the image already overfills the frame, calculate clearance from its actual displayed dimensions.

Clearance prevents blank borders; it does not protect the subject from cropping. Compare the beak, base, face, or label at the tightest frame as a separate check. A large zoom may cover every border while cutting through the object you meant to preserve.

Use FFmpeg for a repeatable centered zoom

If you prefer a command, this alternative creates a centered push-in without a generative model or a video-editor project. Install FFmpeg with the libx264 encoder, save your accepted image as source.png, and run the following in that directory. Use an opaque image; composite transparent artwork onto its intended background first.

ffmpeg -i source.png \
  -vf "scale=3840:2160:force_original_aspect_ratio=increase,crop=3840:2160,setsar=1,zoompan=z='1+0.055*min(on/128,1)':x='iw/2-iw/zoom/2':y='ih/2-ih/zoom/2':d=150:s=1920x1080:fps=30,format=yuv420p" \
  -frames:v 150 -an -c:v libx264 -crf 18 \
  -movflags +faststart motion.mp4

The image covers a 16:9 working canvas, then the centered zoom reaches 1.055 at frame 128 and holds to the end. This version uses linear motion and omits the archived example's diagonal drift. The larger working canvas gives finer crop-position increments; it adds no source detail. Non-16:9 images are cropped, so inspect that framing before using the recipe on a portrait or wide product.

FFmpeg documents the zoompan filter, including its output-frame counter, duration, size, and frame-rate options. This command specifies 150 frames and 30 fps explicitly. If motion.mp4 already exists, FFmpeg asks before replacing it.

To adjust the exercise, change 0.055 for the zoom amount. For a different duration, change the frame total and the motion-end frame together. At 30 fps, ten seconds needs 300 frames; moving through frame 278 leaves a similar short hold. Changing only the frame limit can truncate the move or leave an unexpectedly long hold.

Inspect the encoded result with ffprobe:

ffprobe -v error -count_frames \
  -show_entries "format=duration:stream=codec_type,codec_name,width,height,avg_frame_rate,nb_read_frames" \
  -of json motion.mp4

For the unmodified command, expect one H.264 video stream, 1920 × 1080 dimensions, a 30/1 average frame rate, 150 decoded frames, five seconds, and no audio stream. Those properties establish the file format and timing. They cannot tell you whether a face, a logo, or a mask looks right.

Keep the subject still while the background moves

When you want atmosphere around an unchanged subject, separate the subject from the area that moves. Adobe's masking and rotoscoping documentation describes isolating foreground and background for separate treatment. For a single still, you can begin with a static mask.

1. Duplicate the accepted still. On the upper copy, mask the entire subject, including details such as the base or a contact shadow that should stay attached.
2. Prepare a separate lower background. If you move an unaltered duplicate beneath the mask, its original subject can emerge as a second outline. Remove that background copy of the subject or use a suitable clean background.
3. Keep the foreground subject unchanged. Animate the lower background, add a particle layer, or place a separately generated background video beneath it. If you then want a camera zoom over the whole shot, group both layers and animate the group so the cutout and its surroundings move together.
4. Check the mask against both a light and a dark temporary background. Inspect hair, feathers, handles, gaps between limbs, and translucent edges at full size.
5. Play the complete shot and look for sliding shadows, duplicated features, halos, or background motion that appears to pass in front of the subject.

A mask protects only the pixels it includes. Glass, smoke, reflections, loose hair, and semi-transparent fabric contain information from the original background, so a hard cutout can look wrong even when its outline stays stable. Keep the complete-image move for these subjects unless you can build and inspect the composite carefully.

For modest parallax, split foreground and background into flat layers and move them by different amounts. This can suggest depth, but it exposes areas that were hidden behind the subject. Prepare enough background coverage for the largest displacement. A sideways move cannot reveal an accurate unseen side of a face or product from one flat cutout.

When the subject needs new action, use a short generative test

Use generative image-to-video when the requested shot depends on new action. Check that your account offers an image-input video model, your intended aspect ratio, an appropriate duration, and an acceptable generation cost before starting. If a connected integration offers no suitable video model, check the provider's own interface before assuming your account cannot generate video.

Runway recommends motion-focused prompts, and its Gen-4 guide specifically recommends positive phrasing and warns against negative prompts. Keep the visual checklist outside the prompt. Pixel offsets and exact scale percentages belong in an editor where you can set them directly. The following prompts are suggestions, not tested outputs.

The camera slowly pushes toward the sculpture in one continuous shot. The sculpture remains still.

For environmental movement, a separate starting point is:

Locked camera. The subject remains still while the distant clouds drift slowly across the background.

These are starting prompts, not exact preservation controls. Request one action per initial trial, save the original input, and compare the entire resulting clip. If the scene drifts, remove a movement or try a calmer source image before adding more instructions. A long list of prohibited changes cannot replace output review.

Use a small trial budget you choose in advance. Change one factor between attempts so you can see whether a simpler action or different source helps. If the same face or label keeps changing, switch that element to an unchanged foreground layer or choose a complete-image move. Set aside attempts that change those details, even when the rest of the shot looks good.

For a requested turn or orbit, identify what is hidden in the still. A model must invent those surfaces. When an unseen product side needs to be accurate, supply additional accepted views through a workflow that supports them, or use a real 3D asset. Reframe the shot if those inputs are unavailable.

If the beginning is usable but the subject changes near the end, you can trim the clip before the first bad frame after inspecting everything you keep. To fill the remaining time, hold a last accepted frame or cut to another shot. Avoid extending from a distorted final frame, which makes that altered view the next clip's starting point.

Generate and inspect a first trial in Runway

Runway's Gen-4.5 instructions document the following workflow. This is guidance from the product documentation; the hummingbird sample was made with image keyframes.

1. Open Gen-4.5 in Apps View, or select Video in Tool Mode and choose Gen-4.5.
2. Drag the accepted still into the prompt window. Check that the input preview contains the full face, label, and other details you need.
3. Enter one motion prompt. For the first trial, choose a five-second shot and keep the aspect ratio matched to your prepared image.
4. Check duration, frame rate, and cost before selecting Generate. If you change the aspect ratio, inspect the new crop first.
5. Download the completed output before upscaling or extending it. Compare it with the source and inspect any moment where a face turns, limbs overlap, or a label gets smaller.

The published Gen-4.5 specifications list 720p output with 24 or 25 fps. The 1080p, 30 fps settings in this article belong to the editor and FFmpeg exercises. Inspect the generated file's actual properties before editing it.

Adapt the method to the subject

Portraits, hands, and full-body figures

For a portrait that only needs a sense of movement, animate the complete picture and leave the eyes and mouth unchanged. If the brief requires blinking or turning, compare eye spacing, eyelids, teeth, hairline, jewelry, and the jaw through the action. An opening and closing frame can match while the face distorts between them.

With hands or a full-body figure, inspect finger count, joints, limb overlap, and clothing patterns near the fastest movement. Where two limbs overlap, a generated action may expose anatomy absent from the source. Reduce the action or use a different shot when the missing view cannot be supported. A cropped detail can be a better composition than an unstable full-body movement.

Several people or objects in one image

A slow zoom of the whole picture keeps the spacing between subjects intact. If a generated shot moves them independently, check each person or object, including anything that passes behind another subject. Inspect faces after they reappear, count repeated objects, and watch the gaps between limbs. Reducing movement can make the shot easier to check, but one stable face does not establish that the rest of the group stayed accurate.

Products, logos, labels, and small repeated details

Keep accurate labels and product geometry on the original image layer. Check every visible character and count repeated parts such as buttons, holes, spokes, or dots. A product can look plausible while the model changes its name or adds a component.

If you need a moving background behind packaging, mask the whole package, including its label, as one foreground object. Pasting only a crisp logo over a generated package can leave a mismatch between the logo and a bending surface. For a factual product demonstration, a cosmetic animation should not suggest a feature or mechanism that the product does not have.

Illustrations, architecture, and geometric artwork

Watch thin linework, straight edges, repeated windows, and intersections. A slow zoom can make fine lines shimmer during resampling without bending their geometry. Compare the source, the editor view, and the encoded file to locate where the change appears. Try a smaller move or a source prepared for the output size before adding frame interpolation.

A depth map or separate planes can create a stronger spatial effect, but inspect depth boundaries around railings, hair, and object edges. If the technique stretches or bends those boundaries, return to undeformed flat layers. An orbit around an illustrated building also requires decisions about unseen walls and roof surfaces.

Reflective or transparent subjects

A moving camera would normally change reflections and what is visible through glass. A complete-image zoom keeps the existing reflection attached to the picture, which is useful for a subtle graphic move. For a convincing new viewpoint, treat reflections, refraction, and the background as part of the shot's reconstruction. A preserved outline alone is not enough to judge the result.

Loops and longer clips

The five-second example ends closer to the subject than it starts. Repeating it produces a jump. For a gentle loop, animate out and back to the same scale and position, easing into both ends. Compare the last encoded frame with the first and watch several repetitions. Removing an unnecessary duplicated endpoint frame can avoid a brief pause at the seam.

For example, a five-second loop at 30 fps can begin at 100% scale on frame 0, reach 105.5% at frame 75, and return to 100% at the five-second timeline boundary. Use the same position at both ends and export frames 0 through 149, leaving the repeated first frame to begin the next cycle. Check the seam in your player.

Reversing a camera move is suitable for a still-image zoom. Reversing blinking, walking, smoke, or falling water can reveal unnatural action. For a longer sequence, consider several accepted stills with separate small moves, or hold one accepted view longer. Chaining generated clips from their last frames can carry a changed detail into the next shot.

Inspect the exported clip and fix the failure you see

Watch once at normal speed for movement and once while scrubbing for detail. Then compare full-size frames at the start, around the fastest movement, and near the end. The original example's six-frame contact sheet samples frames 0, 30, 60, 90, 120, and 149. It helps you locate crop changes, but six thumbnails cannot establish that every frame is clean.

Extract individual frames from your delivery file using FFmpeg's frame-selection and output options:

ffmpeg -i motion.mp4 \
  -vf "select='eq(n,0)+eq(n,30)+eq(n,60)+eq(n,90)+eq(n,120)+eq(n,149)'" \
  -fps_mode vfr frame-%03d.png

The extraction command uses frame numbers for the 150-frame, 30 fps exercise. For a five-second clip at 24 fps, use frames 0, 24, 48, 72, 96, and 119 instead. At 25 fps, use 0, 25, 50, 75, 100, and 124. Check the decoded frame count if your file has different timing.

For a pure zoom or pan, details move and change apparent size together. Compare the same part of the subject at a similar viewing size before calling a difference "warping." Look for a local change: a dot disappears, a straight rim bends, a finger duplicates, or a label changes independently of the rest of the object. Image-difference scores also react to scaling, compression, and color changes, so a high score alone does not diagnose deformation.

If you increase frame rate, a still-image transform can be rendered again at the new rate. It does not need an optical-flow pass. After changing frame rate or applying video enhancement, compare the new export with the accepted clip. Fix a bent edge or changed face at its source instead of expecting smoother playback to repair it.

Download the example and adapt it to your image

The hummingbird example uses one complete image, enlarged to 105.5% and moved 18 pixels left and 8 pixels up. The five-second MP4 is 1920 × 1080, 30 fps, and silent. It shows a camera-like move; the sculpture itself stays still.

The example kit contains the source, editable HyperFrames project, MP4, and extracted frames. Use the prompt and review sheet to list the details that matter in your own picture. The original example dates to August 27, 2026 and contains no Runway-generated video. Its optional file checker cannot detect visual deformation; technical notes in the kit explain its limits.

Before using your own animation, check the delivery file at its intended display size. Confirm the details you listed, keep the subject inside the crop, and watch the seam if the clip loops. Save the accepted source and project with the export so you can revise the movement without regenerating the subject.