To animate a portrait subtly, start with a clear image whose face already looks right, choose one small action, and keep the viewpoint close to the original. A blink or a slight settling movement asks for less new visual information than speech, a broad smile or a turn into profile. Compare the moving result with the source at the action's peak and when the face returns toward its original pose.
The goal is recognizable identity through a natural change, not identical facial pixels. Eyelids should change during a blink; the eyes should not emerge with a different shape or spacing. A small turn changes perspective; it should not leave the person with a new jaw, hairline or pair of glasses.
Dreamina offers image-led video creation that can be used for a restrained portrait brief where the selected model accepts the image. The workflow below is documentation-based guidance, with illustrative prompts and review examples. It does not report a completed real-person animation test.
- Decide what kind of likeness the project needs
- Choose a source that does not hide the hardest details
- Set a small motion budget
- Record identity anchors before writing the prompt
- Build one restrained Dreamina attempt
- Review the face at the moments most likely to expose change
- Fix the visible failure, not every part of the prompt
- Use the original portrait when facial reconstruction is unacceptable
Decide what kind of likeness the project needs
A stylized character, a casual moving portrait and a faithful depiction of a specific person have different acceptance standards. Set yours before generating.
For a real-person portrait, identify what makes this photograph recognizable: the face outline, hairline, eye and brow relationship, nose profile, characteristic asymmetry, eyewear or other visible features. You do not need a long anatomical description. You need a short visual reference against which to reject an attractive but different face.
Also decide whether movement inside the face is necessary. A slow approach toward an unchanged photograph may be enough for an introductory title or a memory sequence. A person blinking requires generated facial movement. A talking performance introduces mouth shapes, teeth and timing; it is a different task from a subtle portrait animation.
Use an image and intended depiction you are authorized to use. Check whether the selected model accepts that kind of realistic portrait before planning around it. Model and account restrictions can differ; changing the prompt is not a reason to bypass a rejected input.
For a faithful likeness, the subject's own review can be especially useful. Someone may recognize a changed expression or personal feature that a general reviewer misses. A visually pleasing clip alone does not establish that the portrayal is acceptable.
Choose a source that does not hide the hardest details
The useful source is not necessarily the largest image. It is the one that establishes the features your movement will expose.
A relaxed, near-frontal portrait is a practical starting choice for a small blink or breathing movement. Check that the eyes are distinguishable, the mouth is not already motion-blurred, and the face is large enough to inspect. Leave room around the head and shoulders so a small adjustment does not immediately meet the frame boundary.
For a slight turn, choose a source reasonably close to the intended angle. A frontal image cannot establish the actual hidden side of an ear, hairstyle or face. A side-facing portrait is not a reliable source for inventing an exact frontal likeness simply because the prompt asks for identity preservation.
Inspect these source conditions:
Runway's current image-to-video documentation warns that artifacts such as blurry faces can be intensified in video. It also identifies implied movement in the source as a possible conflict with the requested action. These are useful source checks, not portrait-test results for Dreamina. Runway image-to-video prompting guide.
If you repair or enhance the still, approve it again before animation. Compare the eyes, nose, mouth, skin detail and face outline with the untouched source. A sharpened but altered face is a new problem, not a better identity reference.
The source-image preparation guide covers the broader crop, missing-information and input-role checks. Here, prioritize features that could change the person's likeness.
Set a small motion budget
A motion budget is a plain description of what you will allow to change. It is a planning tool, not a model setting or a proven numeric threshold.
For a first subtle portrait, choose one action and keep other demands quiet:
These are candidate briefs, not an order of experimentally measured reliability. Choose the action your source can support and your viewer needs to notice.
A blink is not automatically harmless: a frame can alter around the eye as it closes. Breathing is not automatically invisible: an exaggerated shoulder movement can look unnatural. The point of a smaller budget is to make the intended change and any unwanted change easier to distinguish.
Avoid combining a turn, smile, blink, hair sweep and camera orbit in the first attempt. If something looks wrong, that combination gives you too many simultaneous changes to diagnose. A new view also makes direct comparison with the original harder.
Record identity anchors before writing the prompt
Keep the approved source available while reviewing. Choose several anchors distributed across the face and its surroundings, rather than relying on one striking feature.
For an illustrative portrait with glasses and side-parted hair, the anchors might be:
- The overall face silhouette and chin shape.
- The hair part and visible hairline.
- The shape and relative position of the glasses frames.
- The relationship between eyes, brows and nose.
- The resting mouth shape and visible characteristic asymmetry.
- The collar and neckline where the head meets the body.
Treat these as visual comparison points, not biometric measurements or a universal identity score. Exact screen coordinates change when the head moves. What matters is whether the features remain coherent with the same person and the permitted action.
Original teaching schematic, not a real portrait or software result. Compare stable identity cues at comparable poses while allowing the requested expression to change.
This preparation helps you notice gradual drift. Looking only at the generated opening and ending can make both feel individually plausible while missing a moment where the face becomes someone else.
Build one restrained Dreamina attempt
Dreamina's current Seedance 2.5 guide documents AI Video > Seedance 2.5 > Omni reference, with + to add references and @ tags to identify their roles. Use that route where the model accepts your image and the controls are available. A general reference guides appearance; it is not the same input role as a dedicated starting frame. Dreamina Seedance 2.5 guide.
Prepare the attempt as follows:
- 1
- Upload the approved portrait and check that the preview contains the intended crop. 2
- Identify the reference as the source of the person's appearance, clothing and setting. 3
- Request the single action selected in your motion budget. 4
- Keep the camera and light steady unless changing them is essential. 5
- Choose a short available duration that can contain that action, and read the displayed credit requirement. 6
- Save the output and the exact instruction so a revision can address one observed problem.
If preserving the exact opening composition is essential and the selected route offers a first-frame slot, use that intended slot. Do not assume that a general-reference tag creates first-frame control.
Example: a blink without a new performance
The following is an illustrative, untested prompt for a portrait that already has a relaxed closed-mouth expression:
Use @Image1 as the reference for the person's appearance, clothing and setting. The person makes one natural blink, then keeps the same relaxed expression. Head and shoulders remain settled. The camera stays fixed and the light stays even. Keep the person's recognizable facial features, hairline and glasses consistent with the reference. One continuous shot.
Replace the reference tag with the label assigned in your workspace. Do not add glasses to the prompt if the source has none; use the features actually present.
If the project needs a tiny head adjustment instead, replace the blink action with that one action. Keep the rest of the brief stable. “Turn into profile and look back” is a substantially different demand from a slight adjustment, even if both take five seconds.
Example: a restrained smile
For a source whose resting mouth is closed, an alternative proposed action is:
The person gradually forms a small closed-mouth smile while maintaining the same head direction. The camera and light remain steady. Preserve the person's recognizable appearance and hairstyle.
This avoids requiring newly visible teeth. It does not guarantee that the mouth will stay closed or that identity will survive. Inspect the returned movement rather than treating the phrase as a protective setting.
Review the face at the moments most likely to expose change
Watch the saved clip at normal speed first. Does the action feel restrained? Does the person remain recognizable without pausing? A clip with acceptable isolated frames can still contain a distracting transition.
Then pause at the opening, the hardest part of the action, and the settling or return. Check additional frames around any suspicious change. Use the full-clip inspection guide when you need a systematic timestamp record; a few checkpoints cannot certify every intervening frame.
For a blink, inspect the eyelids just before closure, during closure and as the eyes reopen. For a head adjustment, inspect the furthest orientation and the return toward the source pose. For a smile, examine the mouth and cheeks as the expression develops, not only the final smile.
Separate expected change from identity drift
A cheek becomes narrower in screen space as the head turns. An eyelid changes shape while closing. Light reflected in glasses can shift with orientation. Those changes can belong to the requested movement.
More concerning signs include a glasses bridge changing construction, a hair part switching sides, an eye returning with a different outline, or a jaw remaining noticeably different after the head returns to a comparable pose. A new tooth pattern appearing in an unintended open mouth is also a reason to reconsider the output.
Compare similar poses where possible. If the person is larger in one frame because the camera approached, account for that scale difference before judging proportions. If the angle never returns, do not claim an exact matched-view identity check was possible.
Reject a clip that looks like a different person even if it is flattering. A likeness requirement is not satisfied by improving the portrait's general attractiveness.
Fix the visible failure, not every part of the prompt
Write down what changed and when it first became noticeable. “Bad face” gives little direction. “The glasses bridge changes as the eyes reopen” identifies a feature and an action phase.
If the opening already differs, check the input role, source crop and any preceding image edit. If the face changes mainly during a turn, reduce the turn or replace it with a less revealing action. If the mouth changes during an added smile, remove that expression before adding more camera constraints.
These are proposed revisions based on the observed symptom, not diagnoses of the model's internal cause. Test one changed variable and assess the new result against the same anchors. A higher output-quality setting does not itself establish a corrected identity.
If the action becomes so small that no viewer notices it, decide whether that is acceptable for the project. The subtle-motion guide helps distinguish intentionally quiet movement from an unclear or unobservable result. Increasing every motion at once may solve visibility while creating a new likeness problem.
Use the original portrait when facial reconstruction is unacceptable
If the person must remain visually exact and internal face movement is optional, build a conventional shot from the approved still. A restrained flat-image zoom moves the photograph without generating a different face. It does not create a blink, breathing or a new viewpoint.
A documented Premiere route is to import the portrait, place it on a sequence matching the intended delivery shape and extend it to the needed duration. Select the image clip, open Window > Effect Controls > Motion, and establish a framing that keeps the complete head visible. Adobe documents the still-image preparation and transform controls. Prepare still images, Motion controls.
To add a gentle approach, enable Scale animation with the stopwatch at the start, then set a slightly larger value near the end. Keep rotation and position fixed unless the layout requires a deliberate change. Preview the tightest crop to make sure it does not cut off the hair or chin. This is a proposed finishing method, not a rendered portrait test. Premiere keyframe instructions.
Export the sequence and inspect the actual file at its delivery size. Check sharpness, crop and the person's appearance after compression. If genuine facial action is essential, an appropriate recorded performance may be a better source than continuing to regenerate a likeness that fails review. Premiere export instructions.
Approve a portrait animation when the person remains recognizable throughout the allowed movement and the action serves its purpose. Keep the original, exact prompt, selected settings and accepted clip together. That gives the next revision a trustworthy reference instead of asking a changed face to become the new standard.