AI video character consistency: plan the person across the shots.
Start with an approved reference, keep a short record of the character's defining details, and review each shot before joining the clips. Repeating a description can communicate your intention, but it does not guarantee the same face, costume, or prop in every generation.
1. Check what your reference input actually controls
First-frame image-to-video and subject-reference generation serve different purposes. A starting image establishes a particular composition. A subject reference asks a supported workflow to carry an identity into another scene. Confirm which mode your selected service and model offer before preparing the inputs.
| Workflow | Prepare | Check before motion |
|---|---|---|
| Approved still → image-to-video | A starting frame for each shot. | The character and composition already look right. |
| Subject references → supported video mode | The subject images that mode accepts. | The selected model supports those reference inputs. |
| Text-only generation | A stable description with the new shot action. | Each result still needs an identity review; text is not an identity lock. |
Runway's Gen-4 Image References guide describes creating still images from character or scene references, then taking a chosen output into a video model. For that workflow, approve the new still before animating it. The image-reference feature itself is not a guarantee of identity throughout a video.
Google's reference-to-video documentation describes a subject-reference mode for selected Veo models, using up to three images of one person, character, or product. Check the supported model and mode in your own interface. Uploading a picture to an unrelated image field does not necessarily activate this capability.
Use references you have permission to use. Choose images where the face, silhouette, and important costume details are visible. If an additional angle is accepted, compare it with the original first: inconsistent source images make it harder to decide which appearance you want.
2. Separate identity from this shot's action
A character card is a manual record for your own decisions. Keep it short enough to compare with a frame. Record the approved files and distinctive details, then write the setting and movement separately. A new action should not silently introduce a new hairstyle or jacket.
Reference files: [approved filenames]. Identity: [face landmarks, hairstyle, silhouette]. Costume: [clothing, colors, distinctive detail]. Props: [object and where it is carried]. Visual treatment: [style]. Story state: [what has already happened]. Allowed changes this shot: [pose, expression, location]. Review notes: [accept or revise, with reason].
For example, an original fictional adult illustrated courier might have short dark curls, a navy jacket with a round brass collar pin, and a red satchel carried at the character's left hip. Those are visible checkpoints. “The same charming hero” is much harder to evaluate.
Keep the card with your shot notes even if the video prompt is brief. Runway's image-to-video prompting guide recommends concentrating on motion when the input image already establishes the appearance. You do not have to paste the whole card into every motion prompt. Use the provider's reference controls and review the result.
3. Keep a two-shot continuity ledger
The example below is an original, unrendered planning exercise, not a report of tested model output. Imagine the courier pausing at a station sign, then walking toward the platform. Prepare two approved starting stills using the same character references where your workflow supports them. Compare the stills side by side before generating motion.
| Shot | Action and frame | Continuity check |
|---|---|---|
| A: pause | Medium view; courier looks toward the station sign. | Face, collar pin and satchel match the approved card. End with the courier ready to leave. |
| B: depart | Wider view; courier walks toward the platform. | The same outfit and satchel remain visible. The new starting pose fits the previous ending. |
Motion brief A
The courier turns their head slowly toward the station sign and pauses. The camera holds a fixed medium view. The shot ends with the courier looking toward the platform.
Motion brief B
The courier takes a few steady steps toward the platform. The camera holds a fixed wider view. The shot ends with the courier still fully visible and the satchel moving gently with their steps.
These prompts assume the approved starting images already show the intended character and scene. They describe the actions to try; the reference filenames in your notes are not universal model syntax. Use your service's actual upload or reference controls.
Record the model, reference files, prompt, duration, and relevant settings for each attempt. Keep them unchanged while diagnosing a particular mismatch. If you need a different angle, approve that starting frame against the original card rather than treating a new generation as the new definition of the character.
4. Diagnose the stage where the character changes
- The starting still already differs: revise the reference setup or still before spending another attempt on motion.
- The face changes during the clip: inspect the turn or movement where it happens. Try a simpler action or a smaller change of view, then review again.
- The costume or prop changes: compare its shape, position, and attachments with the card. Preserve an approved version rather than accepting drift as the next reference.
- Both clips look right alone but the cut feels wrong: compare the last usable frame of A with the first usable frame of B. Check pose, travel direction, location, and story state.
Watch each entire clip, then inspect its opening, middle, ending, and any moment with an obvious change. Compare details relative to the character: a bag can move to the other side of the screen when the camera angle changes without moving to the other hip. Perspective, expression, and lighting can also change how a face looks; decide whether the result still reads as the intended person.
Choose a framing that leaves room for the action with the aspect-ratio guide, and use the camera movement guide when a new viewpoint introduces too much variation. Change one relevant instruction at a time and keep a brief acceptance note for each result.
You can edit and save a motion brief in Shotmelo's image-to-video prompt planner, then export the text for your chosen video service. Keep the character card in your own notes or editable draft. Shotmelo's current preview does not render clips, analyze references automatically, or lock a character's identity.
Common consistency questions
Does using the same seed keep the same character?
Do not treat a seed as a character ID. Its behavior depends on the model and other inputs. Record it when available, but use the supported reference workflow and inspect identity when the prompt or scene changes.
Can I use the final frame as the next starting image?
Some image-to-video workflows support this approach. Approve the frame first: a distorted face or missing satchel would be carried into the next input. A continuation from that frame also serves a different purpose from a new shot with a deliberately different composition.
Can text alone guarantee character consistency?
No. A stable description is useful for planning, but acceptance depends on the generated frames. If a shot fails your identity or continuity checks, revise or leave it out of the sequence.