Create one approved identity, then test it in the angles, expressions, and interactions your drama actually needs. Character consistency requires control over references, scene appearance, and motion; repeating a character’s name or reusing a seed is not a substitute for those controls.

Current tools accept more kinds of references, but that creates a new problem: conflicting inputs. A face image, costume image, location image, and motion clip may all influence the same output. The useful skill is assigning each input a job and detecting when the wrong input takes control.

Design the person at the scale viewers will recognize

For an original fictional detective, choose a few durable features: a broad jaw, a recognizable hair silhouette, a slightly asymmetrical fringe, and a charcoal coat with a light collar. Avoid relying on a tiny ring or subtle eye color to identify the lead in a vertical wide shot.

Now separate the specification:

Layer

Detective example

What may change

Identity

Face structure, hairline, apparent age, proportions

Only through an intentional story decision

Scene appearance

Charcoal coat, dry hair, no visible injury

Clothing, weathering, disguise, or chronology

Shot performance

Turns toward a ringing phone

Action, expression, gaze, framing

This separation is especially important when generating alternate costumes. A prompt saying “a new detective in a red jacket” can invite a new person. An asset connection plus a specific costume change asks for a different operation.

Do not define identity through a famous actor’s likeness unless you have the necessary authorization. A distinctive original design is also easier to manage as a series property than an unofficial imitation of a recognizable person.

Give each reference a role

Prepare a reference manifest before uploading multiple files:

CHAR_01_FRONT: identity and facial proportions
CHAR_01_PROFILE: profile structure, same person
LOOK_02: costume only
SET_OFFICE_A: environment and doorway placement
MOVE_TURN_A: motion intention only, if supported

These are planning labels, not a universal model syntax. In the provider’s prompt or controls, specify which input supplies the person, clothing, setting, and motion.

Runway’s Gen-4 Image References documentation allows up to three active references and recommends clear, neutral, evenly lit source images. Those instructions concern that image-reference workflow; they do not establish limits for every video model.

If a costume reference also contains a different face, crop or prepare it so its intended contribution is clear where the tool permits. Otherwise the generator may combine the garment and the person. Adding still more face references can worsen the ambiguity instead of resolving it.

Do not confuse identity references with start frames

An identity reference tells a system which person to preserve. A start frame also establishes a particular composition and physical state. The distinction matters when the detective must begin with a phone in the left hand rather than merely appear somewhere in an office.

A useful two-stage route is:

  1. Generate or prepare a still containing the correct character, costume, framing, and prop position.
  2. Inspect that still against the scene requirements.
  3. Use it as the starting image for a supported image-to-video route.
  4. Request only the required movement.

    This reduces the number of decisions the video model must invent at once. It does not guarantee that the face or prop survives the motion.

    Conversely, attaching a beautiful portrait as an identity reference does not promise that the first frame will reproduce its pose. Choose the input mode according to whether you need the person, the composition, or both.

    Run a stress test that matches the script

    Do not approve the lead after one attractive frontal portrait. Use a compact test set:

    Condition

    Failure to look for

    First controlled change

    Three-quarter turn to profile

    Nose or jaw becomes another person

    Verify the profile reference before increasing motion

    Quiet speech in close-up

    Face changes with mouth movement

    Compare a simpler delivery or a different speech route

    Hand crosses the face

    Identity reconstructs incorrectly after occlusion

    Simplify blocking or cut before the occlusion

    Two characters share the frame

    Faces or wardrobe blend

    Clarify input roles and test separate coverage

    Low-key lighting

    Age and structure appear to change

    Compare under neutral light to isolate the cause

    The test should include the hardest recurring condition, not every possible pose. If the series is mostly desk conversations, prioritize dialogue and eyelines. If the lead often wears a helmet, test recognition with the helmet before producing an elaborate expression library.

    Record the result as a specific observation: “jaw changes during the right turn,” not “model is bad at faces.” That observation tells you what to repair.

    Diagnose where drift first appears

    Compare the approved reference, the opening frame, a middle frame, and the final frame.

    • Wrong immediately: investigate the selected identity, reference conflict, or initial composition.
    • Correct at first, wrong during motion: investigate occlusion, movement complexity, and the video route.
    • Correct within each shot, inconsistent between shots: investigate costume state, lighting, framing, and which reference version each shot used.

      Do not turn a distorted output into a new reference simply because it is the latest file. That can compound the error across later generations. Return to the approved identity unless a change is intentional.

      When the problem is confined to one difficult action, consider different coverage. A reaction plus a prop insert may preserve the scene more reliably than an extended two-person hand exchange. The decision is editorial: the audience must still understand the complete action.

      Track wardrobe and injuries as story states

      Name looks by scene or story event, such as LOOK_DRY_COAT and LOOK_AFTER_RAIN. Keep the base identity intact.

      For each shot, record the visible costume, held object, injury state, and start/end positions. A scar introduced in episode three must not appear in episode one merely because its later reference is more polished.

      ElserStudio’s canonical World assets can provide a shared reference layer for the production. Use the documented asset and shot connections rather than relying on filenames in unrelated generation sessions. Review the current interface in the official guide.

      Keep experimental outputs separate from accepted identities. A contact sheet is useful only if someone can tell which version is canonical and which is an unapproved variation.

      Decide what consistency is worth correcting

      Not every difference has equal significance. A face swap, uniform change, or vanishing evidence can break the story. A slight background texture change may not.

      Evaluate at normal viewing speed and phone size, then inspect critical frames. A paused-frame artifact in a face or important text can still matter; a tiny incidental variation may not justify discarding a good performance.

      The best character workflow makes intended change possible while accidental change is traceable. It is not a promise that every frame will be identical.

      FAQ

      Can I make siblings resemble one another without blending them?

      Approve two distinct identities and specify shared family traits deliberately. Avoid using one unlabelled face reference as the default for both speakers.

      How should I introduce a disguise that fools another character?

      Create an intentional alternate look and decide which identity features the audience should still recognize. What viewers know can differ from what another character notices; accidental facial drift should not replace that planned reveal.

      Can I switch providers halfway through a series?

      Yes, but test the existing cast under the new route. Provider controls and reference interpretation differ; matching filenames do not establish matching results.