This chapter covers the identity layer of an AI production: how to decide what a character looks like, freeze that decision into a reference artifact, and then reference it — rather than re-describe it — in every shot that follows. It walks from casting and face design through the two routes to a locked lead (a trained personal likeness vs. a generated character sheet), out to non-lead characters and recurring props, and finishes with the naming, folder, and provenance conventions that make dozens or hundreds of generations findable and reusable.
Each generation is an independent event. The model has no memory of the character it drew ten minutes ago, so two prompts describing "the same" person in the same words will produce two different people — close enough to pass in isolation, obviously different once cut together. The failure mode is specific and recognizable: stitched-together AI video clips reveal the same actor with drifting details across shots — different hair or eye color, or a birthmark that jumps position — because nothing locked the character's appearance between generations.
The response running through this whole chapter is the same in every variant: stop re-describing the character in prose, and instead create one durable reference artifact that later prompts point at. A text description is re-interpreted every time; an image is not. That is why Three-Panel Character Sheet exists, why props get their own sheets, and why Higgsfield's Higgsfield @-Element Asset System lets you save a finished reference image as a named element and invoke it by name instead of rewriting its description into each new prompt.
This is the identity half of consistency, not the whole of it. Where a character stands, where the light comes from, and whether the geography of a location holds from shot to shot are separate problems that no reference sheet solves — those belong to Holding Continuity Across Independently Generated Shots. This chapter is only about making sure the thing in frame is recognizably the same thing.
Locking the wrong character is worse than not locking one, because every subsequent shot inherits the mistake. So the first step is casting, and casting means auditions: Iterative Casting for Character Fit is the practice of regenerating a character concept repeatedly until a version's face and energy convincingly read as the intended lead, rather than accepting the first plausible result and trying to direct around a wrong choice later. The workflow in the source video went through roughly three iterations before accepting a take as cast. Three is not a rule — the point is that early generations are treated as candidates to reject, not as output to salvage.
Casting is distinct from directing. Finding the right character is this chapter's problem; getting the right performance out of a character who is already cast is a separate craft, covered in Directing Performance, Dialogue, and Sound. Confusing the two produces a familiar waste: endless performance notes aimed at a face that was never right.
There is also a deliberate design move for the face itself. Deliberate Imperfection as an AI Character Realism Technique treats freckles, birthmarks, scars, and asymmetries as first-class controls rather than defects — in Higgsfield's AI Influencer Studio these imperfection fields sit alongside ethnicity, skin tone, eye color, and age as ordinary parameters. The reasoning generalizes past any one tool: symmetrical, blemish-free renders are a common tell of AI generation, and deliberately breaking that symmetry is a counter-move available wherever a tool exposes fine-grained skin and feature controls. Note the coupling with the previous section — an asymmetric detail like a birthmark is exactly the kind of feature that drifts between shots, so a character built this way makes locking the reference more necessary, not less.
Once a character is cast, the material describes two ways of making that identity durable, and they start from opposite ends.
The first is training a likeness from real photographs. Soul ID: Personal Likeness Training from Reference Photos is Higgsfield's feature for this: supply roughly 20 reference photos of a real person with varied angle, expression, and lighting — not a single headshot — and the trained identity can afterward be dropped into arbitrary scene generations. Functionally, the creator becomes a reusable cast member they play themselves. This is a different mechanism from @-mention assets: nothing is being referenced from a library, a new likeness model is being trained from scratch off personal photos. It is also the other route to "doesn't look like generic AI" — realism arrives because the source was a real face, not because imperfections were hand-designed into a synthetic one.
The second is the Three-Panel Character Sheet: a reference asset consisting of three generated panels — a close-up, a full-body front view, and a full-body rear view — generated once when a lead is introduced, before the character appears in any scene. The three panels exist because a close-up alone under-specifies the character. The close-up carries face detail, the front view carries proportions and outfit, and the rear view carries back, hair, and silhouette — so later generations don't drift on details a single reference never captured. The sheet is generated immediately after casting settles, and then feeds every later generation of that character, in any model.
The two routes compose rather than compete. A sheet can be generated directly from a set of source images — including a trained likeness — so that visual consistency is locked before any shot generation begins, and the sheet becomes the reference all subsequent shots are checked against. Higgsfield's short-film workflow does exactly this: train a personal likeness via Soul ID in HeyGen, run the resulting reference images through Nano Banana Pro to generate a character sheet that locks face, proportions, and styling across angles, then attach or reference that sheet in every subsequent shot prompt for the film. A separate source names the same tool (as "Nana Banana Pro") for building the reference set from multiple angles before the character is uploaded into a video tool. The tool names are incidental; the shape is not — generate the same character from several angles once, then treat that set as the source of truth.
Not everything on screen deserves a full likeness lock, and not everything that needs locking is a person.
For background characters, NPC Character Sheet for Background Character Consistency defines a recurring role rather than an individual. Instead of pinning one person's exact likeness across poses, an NPC sheet specifies age range, general appearance, uniform, and lighting directives for the type — so background characters stay consistent across appearances without needing to be the same modeled person every time. The source's example is a toll-booth controller: the sheet fixed the age range, look, uniform, and lighting for that recurring role before any scene needing one was generated. This is a deliberate reduction in fidelity. A viewer notices when the lead's eye color changes; they notice when the toll booth guy's uniform changes, but not when his cheekbones do.
For objects, Prop Sheet for Cross-Style Object Consistency handles the case where the same physical thing has to survive across scenes rendered in radically different art styles. A prop sheet is a reference image generated for a recurring non-character object — the source's case is a watch — showing a material breakdown, internal parts or mechanism, and multiple angles. It is generated once from the finalized hero image, via a Claude-mediated prompt step and Higgsfield's cinematic image or animation tools; that prompt-generation and tool-chaining layer belongs to Workflow Tooling: Node-Based Pipelines, Claude, and Multi-Model Chains. Once created, the sheet is fed as a style reference into every later scene's generation alongside the scene prompt and the previous scene's rendered video, so the object holds even when the surrounding style changes completely — the same watch whether the scene is a French graphic novel, a manga panel, or a claymation short.
The operative rule the prop-sheet body states plainly: when a project reuses one prop across scenes spanning multiple visual styles, generate a dedicated prop sheet early and treat it as a permanent style-reference input for every subsequent scene, rather than re-describing the object in each new prompt. That is the same instinct as the character sheet, applied to a thing instead of a person.
A locked identity pays off twice — once by holding steady, and once by making variation cheap.
When the story needs the lead to appear differently later — a disguise, a costume change, a damaged or worn state — the move is not to re-run character creation. Character Outfit/State Variation Generation means prompting from the locked reference asset with only the delta described: what changes, and nothing else. Because face, build, and casting are already anchored, the model has a stable base for everything the prompt doesn't mention, and the variation stays recognizably the same character instead of becoming a new one wearing a different coat. Practically, this shortens the prompt as well — the identity has already been paid for.
Character Swap: Scene-First Identity Insertion inverts the usual order of operations entirely. Instead of describing the lighting, mood, and composition you want in text, find an existing photo that already has that look and swap a trained identity into it. The user supplies a Soul-ID-trained character as the source and a separate target photo — one that already contains a person in the desired scene — and the model replaces that person while preserving the target's lighting, mood, composition, and color tones. The consequence is worth sitting with: scene-matching becomes a search problem rather than a prompt-engineering one. Shoot or source reference photos with the right vibe first, then carry the trained identity into them, which sidesteps a lot of the descriptive work that Prompting Foundations: Briefing the Model Like a Director and Camera, Framing, and Color as a Working Vocabulary otherwise ask a prompt to carry.
The source demonstrates Character Swap as one stage of a four-step pipeline built on Higgsfield's Soul 2.0 image model — base generation, Soul ID training, Character Swap, then prop inpainting — for building a consistent AI-influencer or UGC-ad content workflow. That ordering is itself instructive: the identity is trained in the middle, and everything after it is insertion and repair rather than fresh creation.
A reference sheet is only useful if a prompt can reach it. Higgsfield @-Element Asset System is Higgsfield's convention for that: a finalized reference image is saved as a named element with an @-prefixed name — @car sheet, for instance — and from then on it can be referenced by name in later prompts instead of being re-described. That is the whole substrate of asset reuse in the platform, and the thing that turns a folder of good images into a working vocabulary.
Characters get the same treatment. A character produced via a Three-Panel Character Sheet can itself be uploaded and tagged as an @-element in Cinema Studio, given a name and description, and then @-tagged directly inside a shot prompt. This is the concrete mechanism by which a locked identity actually reaches separate shot generations — no more writing out appearance from scratch per shot, no more mismatched hair color or wandering birthmark between cuts. A tagged character also carries into multi-shot generations, staying consistent across the cuts and camera angles produced within a single generation.
Elements don't only come from deliberate upfront asset generation. They can be mined retroactively from generations that already worked: when a sub-element of a shot turns out well — a steering wheel, a piece of furniture — screenshot it and save it as a named element for reuse later. The habit is to review each generation for sub-elements worth keeping rather than only front-loading the asset library at the start of a project. Over a few hundred generations that review pass is what separates a project with a growing vocabulary from one that re-rolls the same object twenty times.
Naming solves reference; it doesn't solve retrieval. Hierarchical Project Folder Structure for AI Generation Projects is the organizational answer: one main folder per scene, subfolders per asset type within it, and a fresh subfolder for every generation iteration. The justification is blunt and empirical — a single commercial project can reach a few hundred generations, past which point folder structure is the only way to locate a specific output. Named elements live inside that structure; the two conventions are complements, not alternatives, and the setup phase of the broader production framework in Workflow Tooling: Node-Based Pipelines, Claude, and Multi-Model Chains is where the structure gets built.
Once more than one person touches the project, the folder becomes a channel for something beyond files. Shared Folders with Generation Provenance for Team Calibration describes attaching full generation provenance — prompt text, model used, and every setting — to each asset, and then sharing that provenance at the folder level rather than file by file. A teammate opening a shared folder sees its assets as if they had generated them themselves: metadata visible, and any asset recreatable with its exact original settings in one click before being edited, animated, or upscaled further.
The value here is structural rather than convenient. Communicating a look normally means messaging settings and screenshotting prompts, which is lossy and manual; folder-level provenance removes that channel entirely by making the artifact carry its own recipe. Higgsfield's Teams rollout calls this "instant calibration." It is the same principle as the @-element system — treat generation inputs as reusable, referenceable assets — scoped to a team instead of a single user's library, and it is what makes a locked character survive a handoff as well as a cut.