A practical reference image workflow for recurring characters in AI video
Your character looks right in the opening shot. In the next clip, her jaw is narrower, her jacket has changed, and she suddenly looks ten years younger. Each video works on its own. Together, they look like a casting mistake.
To keep the same character across multiple AI videos, reuse approved visual references, create and check a starting image for each shot, then animate those images with focused motion prompts. Compare every result with the original character references. A repeated description or fixed seed can support this process, but neither replaces a visual identity reference.
This workflow suits narrative shorts, recurring social videos, music videos, and fictional hosts. It also gives you a way to diagnose what went wrong before spending more credits.
In this guide:
- Why the same prompt produces different characters
- A seven-step workflow for consistent AI video characters
- A three-shot example with reusable prompts
- How to fix face clothing and motion drift
- Tool choices and frequently asked questions
Why the same prompt produces different characters
“A woman with short auburn hair in an olive jacket” describes many possible people. Reusing those words narrows the appearance, but leaves facial structure, age, hairstyle, and proportions open to interpretation. A character name such as Mara helps you organize the project; it does not automatically create a persistent identity in the model.
Two separate problems need attention:
- Consistency between clips: the person in shot two must match the person in shot one.
- Consistency within a clip: the face, body, and clothing must remain stable as the person moves.
A good reference image helps with the first problem. Controlled motion and frame review address the second. A recognizable opening frame cannot guarantee that the face will survive a rapid turn later in the video.
This distinction also explains why “use the same seed” is incomplete advice. Google's reference-to-video documentation describes seed-based reproduction when other parameters remain unchanged. A new scene changes the request. The seed is a reproducibility setting, not a portable character ID.
How to keep the same character across multiple AI videos
Step 1 Define the character separately from the scene
Write a compact character brief with three parts: identity, wardrobe, and visual treatment. Keeping them separate makes intentional changes easier to manage.
For an example character named Mara:
- Identity: adult woman in her early thirties, oval face, brown eyes, chin-length auburn bob with a side part, medium build.
- Wardrobe for this sequence: olive canvas jacket worn open over an ecru crew-neck shirt, dark straight-leg jeans, plain brown ankle boots.
- Visual treatment: naturalistic live action, restrained color grade, soft daylight, realistic skin texture.
Identity persists throughout the story. Wardrobe persists until the script calls for a change. Lighting and framing can vary by shot without becoming a new character design.
Choose a few features that remain visible at the distances you will actually use. A tiny facial mark cannot carry continuity in a wide shot. Hair silhouette, jacket shape, and body proportions remain more useful there. For a cartoon character, record the head-to-body ratio, outline style, and distinctive shapes as well.
Step 2 Build a small approved character reference pack
Choose one clear image as the identity reference. Favor an unobstructed face, readable hair shape, and lighting that reveals facial structure. An extreme expression or heavy shadow makes a poor baseline.
Develop supporting views from that approved design:
- A front or near-front portrait for the face.
- A three-quarter view for angled shots.
- A full-body view for proportions, clothing, and footwear.
These are starting recommendations, not a universal upload requirement. Prepare a profile or rear view when your shot list needs one. Check that every view depicts the same person before accepting it into the pack.
Luma's character consistency guide recommends clean views in separate images and warns that combining angles into one sheet can introduce artifacts. A contact sheet is useful for your own comparison; the generator may work better with individual reference files.
Name the approved assets clearly, such as mara_face_v1.png, mara_threequarter_v1.png, and mara_fullbody_v1.png. Keep experiments elsewhere. The model can only use files you actually upload or select in a supported reference control; typing a filename alone does not attach it.
Step 3 Choose the reference method your tool actually supports
Image inputs perform different jobs. Check the selected model and generation mode before building the sequence.
- A starting-frame image establishes the visible composition at the beginning of the clip. Use it when you want to approve the character, pose, clothes, and setting together before adding motion.
- A subject or character reference supplies appearance guidance while allowing a new composition. Google documents up to three images of one subject for supported Veo reference-to-video models. This is a model-specific capability, not a rule for every generator.
- A performance video plus character reference can guide acting and movement. Luma's Ray3 Modify guide describes applying a character identity to an existing performance, with keyframe controls for modification.
Runway's Gen-4 overview also describes visual references for recurring subjects and coverage from different perspectives. These examples show available approaches; they do not establish a universal winner.
For a first project, start with approved shot images and image-to-video. If you need precise acting, investigate a performance-driven workflow. Do not assume your tool accepts a first frame and additional identity references simultaneously.
Step 4 Approve a still image for every shot
Plan the sequence before generating video. For each shot, write down the framing, starting pose, action, wardrobe, and direction of movement.
Then create a shot-specific still from the approved character references. If your image tool supports editing, change the setting or composition while explicitly preserving identity. Inspect the result before animating it.
Here is a reusable prompt for that image stage:
Use the attached approved Mara references for identity. Preserve her facial structure, apparent age, auburn bob, and body proportions. For this sequence, retain the olive canvas jacket, ecru shirt, dark jeans, and brown boots. Create a medium shot of Mara seated beside a cafe window, holding a closed notebook on the table. Soft daylight, naturalistic live action. Her face is clearly visible.
Adapt the attachment names or reference syntax to your interface. This prompt cannot compensate for conflicting reference images.
View all approved shot images together. Check the face, hair length, jacket cut, and body proportions. Correct a wrong starting image now; animation adds another problem to solve.
Step 5 Animate one manageable action at a time
Once a starting image is approved, let it carry the visible appearance. Use the video prompt mainly to describe movement, timing, and camera behavior. Detailed appearance instructions remain useful for text-only or subject-reference generation, but an image-to-video prompt does not need to redesign everything already visible.
For the cafe image:
Mara lifts her eyes from the notebook toward the window and holds her gaze. Subtle breathing and a natural blink. The camera remains at eye level with a gentle push-in. The lighting stays steady throughout the shot.
Begin with one action. Walking, turning, speaking, picking up an object, and changing camera angle in the same clip makes it harder to identify the cause of a failure.
Choose a generation duration your model supports, then trim to the needed edit length. Shorter usable actions can simplify review, but duration alone does not prevent drift. A brief, fast head turn may be harder than a longer still portrait.
If the face changes during movement, test a smaller turn, a clearer starting angle, or a steadier camera before rewriting the entire character brief.
Step 6 Return to the original references for each new scene
Build scene two and scene three from the approved character pack. Avoid a long chain in which every new scene inherits its identity only from the previous generated scene. Small differences can become the new baseline.
A final frame can help continue an uninterrupted action. It is less useful as the sole identity reference for a completely different setting. Before using one, inspect the face, clothing, hands, and motion blur. Choose a clean frame rather than automatically accepting the last frame in the file.
Keep the same model version and visual treatment while establishing the sequence. If you switch generators, test the approved references on a simple shot first. Different models can interpret the same face or stylization differently.
Step 7 Check the edited sequence against the character pack
Inspect the start, middle, end, and any frames containing a turn, occlusion, expression change, or object interaction. Then watch the clips together at normal speed. Sampling a few frames helps locate problems; it does not replace watching the full result.
Review five things:
- Facial structure, apparent age, hairline, and hair length.
- Body proportions, silhouette, and head size.
- Clothing shape, colors, sleeves, and recurring accessories.
- Stability during motion, especially when the face turns or reappears.
- Continuity across cuts, including screen direction, pose, props, and lighting.
Compare with the original references as well as adjacent clips. Two later shots can match each other while both have drifted away from the approved character.
Reject a clear identity mismatch even when the camera move is beautiful. You can often shorten a clip before a late artifact; a different face throughout the shot calls for a new generation or a targeted edit supported by your tool.
A three shot example you can adapt
Imagine a short story in which Mara leaves a cafe after noticing something outside. The following is a production plan, with suggested edit lengths rather than claimed generation results.
Shot one — the observation, about four seconds. A medium shot shows Mara seated with a closed notebook. Start from the approved cafe still. Animate her looking toward the window with a gentle push-in.
Shot two — the decision, about three seconds. A closer three-quarter image shows the same face and jacket. Create this still from the original character pack, checking it against shot one. Animate a small change in expression as she decides to leave. Keep the camera still.
Shot three — the departure, about five seconds. A full-body image shows Mara outside the cafe with the notebook under her arm. Use the full-body reference to check proportions and boots. Animate two measured steps toward screen right with a slow lateral tracking move.
The third shot does not need to show her standing up, collecting the notebook, opening the door, and crossing the threshold. The cut can imply those actions. That reduces difficult interactions while preserving the story.
For the departure clip, a focused motion prompt could be:
Mara walks two measured steps toward screen right, keeping the closed notebook tucked under her left arm. Her expression remains thoughtful. The camera tracks slowly alongside her at a steady distance. Her movement is natural and the daylight remains consistent.
The reference pack carries identity. The shot images establish appearance and composition. The motion prompts describe what happens. The edit connects the actions.
How to fix common character consistency problems
The face changes between scenes
Check whether each scene used the same approved identity references. Remove contradictory portraits and compare the shot images before generating again. If you used text alone, try a reference-supported workflow.
If the first frame matches but the moving face changes, keep the approved still and simplify the motion. These are different failures and need different fixes.
The outfit changes while the face stays recognizable
Check whether the relevant clothing is visible in the input. A portrait may establish a face while leaving trousers and shoes unspecified. Use a full-body reference to prepare the shot image, and describe the outfit precisely at that stage.
Small jewelry, patterns, and lettering create more details to inspect. Simplify them when they do not matter to the story. Preserve distinctive costume elements when they do.
The character changes after turning around
Supply an approved view closer to the required angle, if the workflow supports it. A front portrait leaves side and rear details unspecified. Alternatively, cut between approved angles rather than requesting a large continuous turn.
Negative prompts do not solve the drift
A long list of defects does not supply missing identity information. Use a clear visual reference and a specific positive motion instruction first.
If your tool has a dedicated negative-prompt field, follow its guidance. Google's Veo prompt guide recommends listing unwanted elements rather than commands such as “don't show.” Do not assume the same negative-prompt behavior exists in every model.
Each clip looks fine but the finished video feels inconsistent
Check framing, screen direction, lighting, and acting intensity. A sudden lens perspective or expression change can make the same character feel different. Matching wardrobe cannot repair a jump from subdued live action to glossy animation.
Use color correction for differences in exposure or white balance. It cannot repair a new jawline or a changed body shape. For a speaking character, reuse the same recorded voice or supported voice preset and check audio continuity separately.
Keep the workflow reusable for the next episode
Save the approved references, character brief, shot images, prompts, model version, and accepted clips together. Record generation settings where they are available. A small production log lets you return to a successful setup instead of reconstructing it from memory.
Before buying more generations, identify the failed stage: reference, shot image, motion, or edit. Change one relevant variable and compare again. Upscale accepted clips after checking the sequence, then review the final exports for new artifacts.
If you use Wery for planning, its Nolan director workflow covers scripts, scenes, storyboards, and production handoffs. Ask for a shot list that names the required character reference and continuity details for each scene. Confirm the chosen video model's reference controls when moving from planning into generation.
Frequently asked questions
Can I keep the same AI character using only text prompts
You can repeat a detailed description, but text leaves many visual decisions open. For a recurring protagonist, use a tool with supported image or character references and check each shot against an approved design.
Does using the same seed guarantee the same character
No. A seed helps reproduce a generation under the conditions documented by the provider. Changing the scene, prompt, reference, settings, or model can change the result. Keep the seed in your log when available, alongside the visual references.
Can the same character wear different clothes
Yes. Keep identity references separate from wardrobe instructions. Prepare and approve a new outfit image before animating it. Compare facial structure and proportions so the clothing change does not quietly become a recasting.
Can I reuse a character across different AI video tools
You can reuse reference assets where the tools accept them, but a character identifier in one platform is not automatically recognized elsewhere. Test a simple shot in the new tool and check the interpretation before producing the next episode.
How do I keep two characters distinct in one scene
Prepare separate reference packs, establish both identities in an approved starting image, and specify positions and actions clearly. Begin with restrained interaction. If faces or outfits blend, try separate close-ups and reaction shots rather than repeatedly regenerating a complicated two-person action.
The next time you make a sequence, approve the character pack and all shot images before animating. That gives every clip a common visual starting point and gives you something concrete to check when a generation goes wrong.
