04.2 Character consistency and embedding-based approaches

01 | THE SAME CHARACTER IN A DIFFERENT SCENE

A campaign shows an explorer beside a lake, walking through a forest and arriving at a station. The settings change, but the audience should recognise the same character.

Character consistency means maintaining recognisable appearance across images. The case study identifies this as important for recurring characters or motifs and suggests embedding-based approaches as a possible aid.

Our explorer campaign is a fictional teaching example. It helps us distinguish identity that should stay stable from scene details that are allowed to change.

02 | DEFINE WHAT MAKES THE CHARACTER RECOGNISABLE

A character can be recognised through a combination of facial proportions, hair, distinctive markings, body shape and signature accessories. No single feature always establishes identity.

For a classroom character, we might specify short curly hair, a round face, a green jacket and a triangular badge. These are proposed design choices, not details in the booklet.

A green jacket alone is weak evidence: many different characters could wear it. Review the combination of identifying features rather than counting one matching colour as success.

03 | IDENTITY IS NOT POSE, EXPRESSION OR BACKGROUND

A consistent character should usually be able to move, smile, turn or appear in a new environment. Consistency does not mean duplicating one image.

Separate stable features from permitted variation. A different pose can be correct; an unexplained change in facial proportions may not be.

A fair evaluation includes varied views and contexts. Generating nearly identical portraits can make a system appear consistent without testing whether it supports useful creative variation.

04 | CHARACTER AND STYLE CAN VARY INDEPENDENTLY

Style describes artistic treatment, such as a flat illustration or photographic rendering. Identity concerns who or what the recurring character looks like.

Two images can share a palette and brush treatment while showing different faces. The same recognisable character could also appear in two different artistic treatments.

Visionary Studios may require both stable identity and a coherent campaign style. Score them separately so that success in one does not hide failure in the other.

05 | WHY REPEATING THE PROMPT MAY NOT BE ENOUGH

A description such as “an explorer in a green jacket” leaves many facial and body details unspecified. Different samples can satisfy those words while depicting different people.

Clearer descriptions can help, but ordinary text guidance does not guarantee exact identity. Changes to pose, lighting or composition can also interact with the resulting appearance.

A fixed seed is not an identity lock. It controls a pseudorandom process under compatible conditions; changing the prompt or system can still change the character.

06 | WHAT IS AN EMBEDDING?

An embedding is a numerical representation of information in a learned representation space. It is often a vector: an ordered list of numbers.

For example, an encoder can map an image into a vector that captures features useful for its training objective. A text encoder can also produce numerical representations of words or descriptions.

The numbers are not normally a simple checklist with one labelled coordinate for hair and another for a jacket. Features can be distributed across dimensions, and what is captured depends on the model and its training.

07 | AN ENCODER CREATES THE REPRESENTATION

An encoder transforms input data into a representation. A reference-image encoder can turn an approved image into information that a compatible generation system can use.

This is a conceptual workflow. It does not imply that every generator accepts arbitrary vectors or that an embedding alone contains a complete generation system.

08 | HOW AN EMBEDDING CAN GUIDE GENERATION

A compatible model or conditioning component can use a reference representation to influence its predictions. Text or other inputs can request a new setting while reference information helps retain important appearance.

The aim is useful preservation, not necessarily exact copying. The generated pose and background can change while identifying features remain recognisable.

The booklet names embedding-based approaches broadly. This lesson explains a possible mechanism rather than claiming that Visionary Studios has chosen one particular implementation.

09 | NOT ALL EMBEDDINGS CAPTURE THE SAME THING

A general image embedding may capture broad subject matter, composition or style. An identity-focused representation is trained to emphasise different features.

A high general similarity score might reflect two green jackets and matching backgrounds rather than the same face. A face-focused representation might miss an important badge or costume requirement.

Choose a representation and evaluation method appropriate to the task. Human and fictional non-human characters may need different checks, and compatibility between components matters.

10 | SUPPORTING METHODS ARE NOT IDENTICAL

Reference conditioning can use an existing image encoder and generation components without ordinary updates to the base model. Other approaches learn a specialised representation from examples, or adapt selected model parameters.

Learning a new embedding is a training process. Fine-tuning changes some or all selected model parameters; merely supplying a reference during inference is a different operation.

Do not describe every character-consistency technique as fine-tuning or every fine-tuning method as “just an embedding”. Their preparation, data and computing requirements differ.

11 | PREPARE REFERENCES CAREFULLY

Use approved references that clearly show the intended features. Several compatible views can help communicate appearance across angles, but poor or conflicting references can create ambiguity.

Separate identity information from incidental details. If every reference has the same background, a system may reproduce unwanted background features or struggle to vary them.

Keep a character sheet describing stable features, allowed changes and suitable views. Investigate permissions and data handling before uploading client material or a person’s likeness to an external service.

12 | PRESERVATION AND FLEXIBILITY CAN COMPETE

Some systems offer controls affecting the influence of reference guidance. Stronger guidance may preserve more reference appearance, but can also constrain pose, scene changes or style.

Weaker guidance may allow more variation while losing identifying features. The behaviour and numerical scales depend on the system.

Test whether the method preserves the right properties. Copying a background very faithfully is not success if the brief requires a new location and the face still changes.

13 | SIMILARITY SCORES ARE EVIDENCE, NOT IDENTITY PROOF

Representations can be compared using measures such as cosine similarity, which compares vector directions. A high score means similarity according to that representation and measure.

It does not establish perfect identity preservation or satisfy every campaign requirement. Lighting, style, background and the encoder’s limitations can influence results.

There is no universal acceptance threshold that works across all encoders and tasks. Validate any proposed threshold with relevant examples and use appropriate visual review alongside it.

14 | TEST A WHOLE CAMPAIGN

Generate the character across varied poses, expressions, backgrounds and lighting conditions that matter to the project. Compare both with approved references and across the generated set.

Record failures as well as selected successes. A campaign with one inconsistent image may still need revision even if the remaining images are strong.

15 | REVIEW A FICTIONAL RESULT

Suppose a generated forest scene preserves the green jacket and badge, but changes the character’s facial proportions. A general image similarity measure reports a high score.

The right judgement is not “the number proves consistency”. The shared clothing or scene treatment may contribute strongly to that score. Inspect the identity features separately.

A revision could investigate better reference views, compatible identity guidance or editing. Record the change and recheck pose, style and the rest of the image, because an identity correction can affect other requirements.

16 | RECOGNISE LIMITATIONS AND RESPONSIBILITIES

Large viewpoint changes, occlusion and unusual lighting can make preservation difficult. A method that works for one character may not transfer equally well to another.

Using representations does not automatically remove privacy, permission or bias concerns. An embedding can retain information about its source; it should not simply be declared anonymous.

For the studio, a useful pilot should consider review effort, rejected images, reference preparation and data handling alongside consistency improvements.

17 | APPLY IT TO VISIONARY STUDIOS

A justified recommendation balances recognisable identity, creative flexibility and workflow cost. It does not promise that one embedding guarantees the same character forever.

18 | DESIGN A CONSISTENCY TRIAL

In your book, define four stable character features and three allowed variations. Propose a comparison between text-only generation and a compatible reference-based approach.

State which references you would prepare, what you would record and how you would judge identity independently of style and background.

Read alongside the character consistency and embedding-based approach discussion in the official 2027 booklet. The explorer design and detailed methods here are teaching extensions.

CHECK YOUR UNDERSTANDING | IDENTITY, VARIATION OR REPRESENTATION?