How to Build an AI Character Asset Library: From One Series to a Franchise

The AI casting guide covers how to build a reference pack that holds character consistency through one series. This guide covers what happens after that series performs well and the production company needs the same character in a sequel, a spin-off, a promotional short, and a brand integration, all produced by different operators across six months.

The frontier in 2026 is no longer episode-to-episode consistency. It is universe-level consistency: the same locked character appearing across a flagship series, a prequel, a promotional short, and a brand integration, identical in all of them, produced by different operators months apart.

A reference pack is not a character asset library. A reference pack is a collection of approved images that a generation operator uses in a session. It lives in a folder. It requires the operator to select and upload the correct images before each generation. It produces good results within a session and requires re-validation across sessions. It does not transfer cleanly between operators. It does not persist when a new series starts production six months later.

A character asset library is a persistent production infrastructure that stores character identities as named, trained, version-controlled assets that any authorized operator can access and deploy without rebuilding the reference from scratch. The difference between those two things is the difference between a production company that can produce a sequel and a production company that has to re-cast its own characters.

Why Reference Packs Fail at Franchise Scale

The reference pack failure mode at franchise scale is predictable and consistent. Session memory drift.

Drift happens when identity is stored in temporary session context rather than persistent assets. Small variations compound across shots and sessions until the character visibly changes. Reference images condition individual generations but do not persist production state, enforce identity across regenerations, or transfer cleanly between team members over a 60-plus episode season.

The specific failure pattern for a franchise: the first series builds a character reference pack of 25 to 30 approved images. The series generates well. The paywall conversion rate confirms that viewers are investing in the character. Six months later, the sequel enters pre-production. The original generation operator is unavailable. A new operator rebuilds the character from the existing reference pack images. The first ten episodes of the sequel look close to the original character but not identical. By episode twenty, the differences have compounded to the point where a viewer completing both series notices that something is off about the character without being able to identify what.

That viewer's perception of the franchise is damaged not by the story quality but by a production infrastructure decision: the character was stored as a reference pack rather than as a persistent locked asset.

The Character Lock vs Reference Pack Distinction

Character lock stores an approved character as a permanent node in a production graph, so every shot in every session references the same enforced identity instead of a remembered one. Ask three questions before trusting any character consistency claim. First, does consistency survive a closed session and a new operator, or only a continuous run? Second, can the same locked character appear in two different series without rebuild? Third, when a shot fails and regenerates, does identity hold automatically or does it depend on the prompt?

The practical difference in a franchise production: a reference pack requires the operator to upload images before each session. A character lock requires the operator to select the character's name from a dropdown. The result is different because the mechanism is different. The dropdown selection retrieves a trained model that enforces the character's specific facial geometry, bone structure, skin tone, and wardrobe characteristics at the generation level. The image upload conditions the model on reference images that it interprets differently in each session.

For a franchise production company, the character lock is not a quality preference. It is an operational requirement. A franchise character that can only be reproduced reliably by the original operator who built the reference pack is not a franchise asset. It is a production dependency.

Stage 1: Building the Foundation Asset for Each Character

The character asset library begins with the foundation asset: a trained character identity that becomes the permanent reference for all subsequent franchise productions.

The Photo Selection for Soul ID Training

Soul ID solves the AI character consistency problem: train it once on 20 or more photos of a person, and it locks their facial features across every generation, regardless of pose, outfit, lighting, or style.

The photo selection criteria for Soul ID training in a vertical drama franchise context are more specific than the general Soul ID guidance because vertical drama's 9:16 close-up frame has specific requirements:

Photo count: 20 photos is the minimum. 30 to 40 photos is the production-grade standard for franchise characters. More photos provide more facial geometry data across more angles and lighting conditions, which produces better generalization when the trained model is deployed in generation conditions that differ from the training photos.

Angle coverage: The training set needs to cover the full range of angles the character will appear in across franchise productions. Front-facing, three-quarter left, three-quarter right, slight upward angle, slight downward angle. Each angle teaches the trained model how the character's specific facial geometry behaves in that viewing position. Missing an angle in training produces inconsistency when that angle is used in generation.

Lighting coverage: Include photos in at least three distinct lighting conditions: warm key from one side, neutral even lighting, and cool or high-contrast dramatic lighting. Vertical drama uses all three lighting conditions across its scene types. A trained model that only has warm key lighting photos will produce inconsistency when deployed in a high-contrast confrontation scene.

Expression coverage: Neutral expression baseline, controlled tension (the suppressed emotion register that vertical drama uses most), warmth approaching vulnerability, and the specific expression the character wears at the paywall episode's button cut. The last category is the most franchise-specific: the expression the character carries at the moment of maximum emotional investment is the expression that the viewer's parasocial memory locks onto. Training the model on this expression ensures the franchise's sequel deploys it correctly.

What to exclude: Higgsfield's Soul ID documentation specifically notes: avoid sunglasses, masks, and extreme expressions. These give the AI incorrect reference data that degrades the trained model's performance in standard generation conditions. Also exclude photos with significant hair variation from the character's established look, heavy makeup that differs from the character's established appearance, or photos taken in environments with strong color casts that would affect the skin tone training.

The Training Process

The Soul ID training process takes approximately 5 minutes once the photos are uploaded. The training analyzes the photos and builds a digital twin capturing the character's specific face shape, skin tone, bone structure, hair characteristics, and expression patterns.

Once training completes, the character receives a name within the production's library system. The name is the production-grade identifier that operators use to deploy the character in generation. It is not the character's story name. It is the library asset identifier: Lead_Female_S01, CEO_Antagonist_S01, Secondary_Male_S01. The naming convention reflects the franchise's production structure rather than its narrative structure, which prevents identifier confusion when the same character appears across multiple series.

The training is completed once per character per major appearance change. A character who appears with the same hair color and general appearance across the first and second series of a franchise shares one trained Soul ID asset. A character who undergoes a significant appearance change between series, a different hair color, a significant age shift, a wardrobe category change, requires a second trained Soul ID asset for the post-change appearance.

Stage 2: The Franchise Asset Library Structure

A franchise character asset library is not a folder of images. It is a structured system that stores trained character identities, approved generation outputs, wardrobe configuration records, behavioral documentation, and version history in a format that supports multi-operator production across multiple series.

Asset Category 1: Trained Identity Models

The core of the library. Each character's Soul ID-trained model, stored with its training date, training photo count, and the generation operator who built it. Version numbers are assigned when the character's appearance changes: Lead_Female_S01_v1 is the character's trained identity for the first series, Lead_Female_S01_v2 is the trained identity for the post-change appearance in the sequel.

Access control: the trained identity models are available to all authorized operators in the production pipeline. An operator working on a brand integration that features the franchise's lead character accesses the same trained model as the operator who generated episode one of the original series. This is the library's core value: the character is available to any authorized operator without rebuild.

Asset Category 2: Approved Generation Output Archives

Every generation output that was approved for use in a produced episode is archived with its episode number, scene type, generation tool, prompt text, and operator ID. The archive serves two functions.

First, approved outputs become the highest-quality reference for subsequent franchise productions. A generation operator producing episode one of the sequel can review the approved outputs from the original series to understand how the trained character model renders in specific scene types and lighting conditions. The approved outputs are more specific than the reference photos because they show the character as the generation tools have already rendered them in the franchise's actual production conditions.

Second, the archive is the quality audit trail. If a franchise production delivers content where the character's appearance is questioned by the platform, the archive allows the production company to trace the specific trained model version, prompt specification, and generation tool that produced the questioned content.

Asset Category 3: Wardrobe Configuration Records

Each wardrobe configuration the character wears across franchise productions is documented with its approved reference images, its episode range notation indicating when the configuration is current in the narrative, and its generation prompt specification for how to deploy the configuration in new generations.

The wardrobe configuration record prevents the most common franchise continuity error: a character wearing the correct general wardrobe type but in the wrong specific configuration for their position in the narrative arc. A controlled alpha who began the first series in full defensive armor, his wardrobe, and progressively revealed vulnerability as his wardrobe became less formal across the arc needs precise wardrobe configuration documentation. A sequel production company that accesses the library without this documentation may deploy the character in the wrong wardrobe configuration for the sequel's narrative position.

Asset Category 4: Behavioral Documentation

The behavioral documentation records the character's specific physical behavioral tells that the franchise audience has come to recognize. Not the character's personality description. The specific physical behaviors that visual drama generation needs to reproduce:

Which hand the character uses to express suppressed anger. The specific stillness that communicates the character is processing something significant. The head movement the character makes before they speak after a long pause. The eye movement pattern that reveals the character is lying.

These behavioral specifics are invisible in reference photos and cannot be reliably specified in a prompt without documentation. They are the parasocial details that franchise audiences carry from one series to the next and that, when correctly reproduced in the sequel, produce the recognition response that franchise loyalty is built on. When incorrectly reproduced or absent, they produce the uncanny sense that the character is close to but not quite the character the viewer knows.

The behavioral documentation is written by the creative director who produced the original series and reviewed against the approved generation output archive before being filed in the library. It is the hardest category to capture and the most franchise-valuable.

Asset Category 5: Version History and Change Log

Every change to any asset category is logged with its date, the operator who made the change, and the reason for the change. The version history allows the production company to understand how the franchise's character assets have evolved across productions and to revert to an earlier version if a change produced an inconsistency that a subsequent production needs to correct.

Stage 3: The Soul ID to Generation Pipeline in Franchise Production

The franchise production pipeline differs from the single-series pipeline in the generation workflow's starting point. Single-series production begins by uploading reference images to the generation session. Franchise production begins by selecting the character's named asset from the library.

Connecting Soul ID to Generation Tools

Train the spokesperson in Soul ID, then generate through Seedance 2.0. The identity applies across every format without rebuilding the reference each time.

The connection workflow for each major generation tool:

Seedance 2.0: Connect the Soul ID trained model to the Seedance 2.0 generation workflow through the character reference input parameter. The trained model's identity persists across all generation sessions that reference it. A new operator generating episode one of the sequel selects the same trained model that the original operator used for episode one of the first series and receives the same character identity in the generation output.

Kling 3.0: Use Kling 3.0's Elements system with the approved reference images from the character asset library's approved generation output archive. The Elements system's @ reference syntax deploys the character in multi-shot sequences. For franchise productions using Kling 3.0 for action-adjacent and multi-shot scenes, the reference images from the approved output archive are more effective reference inputs than the original training photos because they show the character as Kling 3.0 has already rendered them.

Veo 3.1: Upload the character's approved reference images from the library's approved generation output archive as the character reference input. Veo 3.1 does not have a persistent trained model system equivalent to Soul ID, so franchise character consistency in Veo 3.1 depends on the quality of the reference images rather than on a trained model. Select the highest-quality approved outputs from the character's archive in the scene type Veo 3.1 will be generating for the new production.

The Generation Session Protocol for Franchise Characters

Before any generation begins in a franchise production, the generation operator completes a character asset library check: confirm the correct trained model version is active, confirm the correct wardrobe configuration references are loaded for the new production's narrative position, and run a test generation of the character in the new production's standard interior environment.

The test generation is reviewed against the approved output archive from the original series. The review question is not whether the test generation looks good. It is whether the test generation looks like the same character who appeared in the approved outputs from the original series. The franchise audience will make the same comparison. The review catches any version discrepancy or model drift before production generation begins rather than discovering it at the rough cut review.

Stage 4: Character Asset Updates Across Franchise Productions

The character asset library is a living document that evolves with the franchise. Managing updates without introducing inconsistency is the operational challenge that separates well-managed franchise production companies from those that discover character drift at the rough cut stage.

When to Update a Trained Identity Model

A trained Soul ID model is updated when:

The character's appearance changes significantly between productions. A sequel where the character is older, has different hair, or has a different wardrobe category from the original series requires a new trained model for the post-change appearance. The new model is given a new version number and the old model is archived, not deleted.

The original training set is found to produce inconsistency in a specific angle or lighting condition that the new production requires. Additional training photos are added to address the specific gap and the model is retrained.

A generation tool update changes how the trained model is deployed. When Soul ID or its connected generation tools release significant updates, the trained models should be tested against the approved output archive to confirm that the update has not altered the trained model's output characteristics.

When Not to Update a Trained Identity Model

A trained model is not updated because the new production's generation team wants to make minor adjustments to the character's appearance. The trained model represents the franchise character as the audience knows them. Adjusting the trained model because a new operator has aesthetic preferences about the character's appearance is a franchise consistency violation, not a legitimate update.

The franchise character asset library has a designated owner, typically the creative director of the franchise's original series, who approves all updates to the trained identity models. Updates to wardrobe configuration records and behavioral documentation can be made by generation operators with the creative director's review. Updates to trained identity models require the creative director's approval before the new version is deployed in any production.

The Franchise Character as a Licensable Asset

A locked character is a durable asset. It can anchor sequels, spinoffs, and brand partnerships because it can be reproduced exactly, on demand, indefinitely. A session memory character is a recipe that produced someone similar last time. Only one of these is an asset a studio can build IP value on, and as vertical drama attracts brand budgets, reproducible characters become licensable ones.

The commercial implication for vertical drama franchise production companies: a character asset library that contains trained identity models for 3 to 5 franchise characters, with version histories across 2 to 3 franchise productions and approved generation output archives that demonstrate the characters' consistent deployment across those productions, is a demonstrable IP asset.

When a platform commissioning team evaluates a production company's sequel pitch, the character asset library is the evidence that the sequel will reproduce the audience's parasocial connection to the characters that the first series built. A production company that can demonstrate the library to the platform's acquisition team is demonstrating something a production company with only a reference pack cannot: that the sequel's characters are the same characters as the original series, not approximations built from documentation.

Axis AI Studios Perspective

The character asset library is the production infrastructure decision that converts a successful series into a franchise business. The production company that completes a first series successfully and stores the character as a reference pack has built an asset that will require significant reconstruction work if the series performs well enough to warrant a sequel. The production company that completes a first series successfully and stores the character as a trained, named, version-controlled asset in a structured library has built an asset that the sequel can deploy from day one without reconstruction.

The value difference between those two positions compounds across franchise productions. A four-series franchise built on a properly maintained character asset library generates value from each series' character development for every subsequent series in the franchise. A four-series franchise built on reference packs regenerates the same reconstruction investment before each new series.

At Axis AI Studios, the character asset library is built in pre-production of every series, not at the point where a sequel is commissioned. The trained models, the approved output archives, the wardrobe configuration records, and the behavioral documentation are established before episode one generates, not after the paywall conversion rate confirms the series justifies a franchise.

For production companies who want to build franchise-grade character infrastructure from the first series, reach out at business@axisaistudios.com.

Character Asset Library Maintenance Checklist

After each episode batch is approved:

• Archive all approved generation outputs with episode number, scene type, tool, and operator ID

• Flag any outputs where character appearance was outside the approved range for operator review

• Update the wardrobe configuration record if a new configuration appeared in the batch

After each series completes:

• Review the approved output archive for the highest-quality outputs in each scene type

• Update the behavioral documentation with any behavioral specifics that emerged during production

• Confirm the trained model version number that was used for the series and archive it

Before each new franchise production begins:

• Run the character asset library check for each franchise character

• Confirm the correct trained model version is active

• Run test generations against the approved output archive from the most recent prior production

• Distribute the current library documentation to all operators assigned to the new production

FAQ

How Many Photos Are Required for a Production-Grade Soul ID Training?

Twenty photos is the minimum that Soul ID requires. For franchise productions where the trained model will be deployed by multiple operators across multiple series over 12 to 24 months, 30 to 40 photos is the production-grade standard. The additional photos above the 20-photo minimum cover angle and lighting conditions that the minimum training set may not adequately represent, producing better generalization when the trained model is deployed in generation conditions that differ from the training set.

Can a Soul ID Trained Model Be Shared Between Production Teams?

Yes, within a production company's authorized operator network. The trained model is stored in the production company's Higgsfield workspace and accessible to any operator with workspace access. For productions involving multiple production companies, the trained model can be shared by adding the external operator to the workspace with appropriate access permissions. The sharing mechanism is the production company's decision and should be governed by the IP agreements that cover franchise character ownership.

What Happens if a Generation Tool Update Changes How the Trained Model Performs?

Test the trained model against the approved output archive immediately after any significant generation tool update. Run 10 to 15 test generations across the scene types most represented in the approved output archive and compare against those approved outputs. If the update has materially changed the trained model's output, either retrain with additional photos to correct the drift or document the change in the version history so that subsequent franchise productions understand the tool version context for each approved output in the archive.

Further Reading

For the character reference pack that precedes the full asset library described in this post, the AI casting guide for vertical drama covers the reference pack construction workflow and the test session protocol that validates each character before production generation begins.

For the franchise architecture that determines which characters warrant full asset library investment and which are episodic roles that do not, the guide to building a microdrama franchise covers franchise strategy and the character design principles that produce franchise-ready characters.

For the prompt engineering workflows that deploy the trained character assets in Seedance 2.0 and Kling 3.0 generation sessions, the prompt engineering guide for vertical drama generation covers the specific language structures that produce consistent output when connected to a trained Soul ID model.