Higgsfield Soul

Higgsfield Soul ID — Character Consistency

QUICK FACTS

Generated-checked block (scripts/build_index.py verifies anchors). Read the linked sections for full context — these lines are routing aids, not the rules themselves.

• Hard rule: every Soul ID prompt splits into Identity Block (static descriptors only) + Motion Block (temporal/camera only) →
• Don't re-describe the face or core features — only describe what differs from the base character →
• Reference image rules: front or 3/4 angle, even lighting, neutral-to-slight expression, no blur, solo subject →
• Reference generators: Soul 2.0 (fashion-forward), Nano Banana Pro (max sharpness), Seedream 4.5 (style range) →
• Character sheet angles: front face, 3/4, side profile, optional full body + optional embedded prop sheets →
• Two-image floor per character: one clear face + one full body, on neutral grey — never a single image; the grey law's shade and mechanism live once in ../../templates/ad-asset-prep.md →
• The Untouched Base: face pass in close-up, then a looks pass; the base close-up is never run through a model again — changes go on with masks, and a new state is a new named asset →
• Reference plates: biological realism ON, photographic capture behaviour OFF (no key direction, cast shadow, bokeh, flare); the background is a field, not a room; lean prompts once a reference carries identity →
• Prefer the single-prompt 3×2 six-panel sheet (one 16:9 generation) — identity locks better than multi-step assembly →
• Outfit change without identity drift: split-panel sheet — ghost-mannequin outfit LEFT + "face matches input 100%" close-up RIGHT →
• Crowds: a single-character reference makes a clone army — build a multi-character lineup and declare it a "VARIETY reference" →
• Character Anchor Block = 10 per-shot attributes (identity, screen position, depth layer, frame occupancy, orientation, pose, gaze, contact points, state lock, expression) →
• One Soul ID sheet PER character state — 5 transformation stages = 5 distinct sheets; the prompt names the stage →
• Micro-expression presets: 9 core + 10 extended →
• Cinema Studio: identity goes in the @ Element definition, motion goes in the prompt field →
• Soul Cast is Business/Team plans only; 8 parameter categories incl. 12 Archetypes, Budget $10M–$500M, Era 1900s–2020s →
• Soul Cast specs: up to 4K (Character/Location modes) / 2K (General); batch 1 or 10; 0.125 credits per image →
• Use 2–3 reference shots (frontal, 3/4, side); if features drift, use the character sheet as @Image1 →
• Soul Cinema is the default CS 3.0/3.5 image-mode model — single-step, ~0.125 credits/image →
• Studio-feeling output is an intermediate, not a final — re-pass through Soul Cinema with grade, directional lighting, lens character; the result is a derived look frame, never the identity base or a reference plate →
• Skip Soul ID for single shots or when you want maximum creative variation →
• Plasticky face in wide shots: crop the face from a closer-shot panel and replace it in post →
• Casting a REAL person: the sheet alone returns a lookalike — erase the head on the body panels and paste their actual photo into the portrait panel →
• Run the same sheet prompt through 2–3 image models and compare before locking; the deciding axis changes per character →

What Is Soul ID?

Soul ID is Higgsfield's character consistency system. Create a character reference once, then reuse it across unlimited generations — different scenes, angles, lighting, and actions — while the face and core appearance remain consistent.

How It Works

1. Create a Soul ID — Upload a reference image or generate one using Soul 2.0
2. The platform stores the character — assigns it an ID
3. Reference in future prompts — "using Soul ID character [name/reference]"
4. Character stays consistent — face, skin tone, basic features carry across generations

When to Use Soul ID

✅ You're building a multi-shot sequence with the same character ✅ You're creating a short film / story with a recurring protagonist ✅ You're making an AI influencer or brand mascot ✅ You want consistent faces across a product ad campaign ✅ You're doing a multi-scene action sequence and need the hero to look the same

❌ Don't use if you only need one shot — overkill for single generations ❌ Don't use if you want maximum creative variation — consistency limits style range

Prompting With Soul ID

When a Soul ID character is active, your prompt should:

1. Reference the character simply:

The Soul ID character walks through a crowded Tokyo street at night.

2. Describe what changes — not what the character looks like:

The Soul ID character now wears a formal black suit. She stands at a podium,
addressing a conference room. Camera: Dolly In toward her face.
Style: Cinematic, cool corporate lighting, 16:9.

3. You can change clothing, setting, expression:

The Soul ID character is now in a red dress, dancing alone in a ballroom.
Camera: 360 Orbit.
Style: Cinematic, warm golden chandelier light.

Key rule: Don't re-describe the face or core features — the Soul ID handles that. Only describe what is different from the base character.

Identity vs. Motion Separation — Hard Rule

When Soul ID is active, every prompt MUST be split into two blocks. This is the single most important rule for preventing identity drift.

Identity Block — Static descriptors only

Contains: face features, clothing, body type, distinguishing marks, color palette. Does NOT contain: any motion, camera, speed, or temporal language.

Motion Block — Temporal and camera only

Contains: camera movement, action choreography, speed, environmental changes. Does NOT contain: any character appearance repetition.

Before/After Examples

Example 1 — Action scene:

❌ Mixed (bad) — causes identity drift:

A tall woman with green eyes and freckles in a leather jacket sprints through
a warehouse while the camera tracks her and her green eyes flash with determination
and her freckles catch the fluorescent light as she vaults over a railing.

Face morphs mid-clip because the model re-reads face descriptors while processing motion.

✅ Separated (good) — identity stays locked:

Identity Block:

The Soul ID character — tall build, green eyes, light freckles across the nose
and cheeks, wearing a fitted black leather jacket, dark jeans.

Motion Block:

She sprints through a dimly lit warehouse, vaults over a metal railing
without breaking stride.
Camera: Action Run — low behind her, matching pace.
Fluorescent lights flicker overhead.
Style: Cinematic, cold industrial blue, high contrast. 16:9.

Example 2 — Emotional close-up:

❌ Mixed (bad) — face warps during camera move:

A weathered man with deep wrinkles and sad brown eyes wearing a grey wool coat
sits on a park bench as the camera slowly dollies in on his wrinkled face and
sad brown eyes while autumn leaves drift past his grey coat.

✅ Separated (good) — face stays sharp:

Identity Block:

The Soul ID character — grey-haired man, deep wrinkles, warm brown eyes,
wearing a heavy grey wool coat, brown leather gloves.

Motion Block:

He sits on a park bench, hands folded in his lap, staring at the ground.
A single autumn leaf drifts into frame and lands on the bench beside him.
Camera: slow Dolly In toward his face.
Style: Cinematic. Overcast diffused light, muted earth tones. 16:9.

Example 3 — Cinema Studio (@ Elements):

❌ Mixed (bad) — identity in the prompt field:

@Sarah with her dark curly hair and tattoo sleeve walks into the bar.

✅ Separated (good) — identity in the Element, motion in the prompt:

@ Element definition (set in Cinema Studio UI):

@Sarah: dark curly hair, tattoo sleeve on left arm, wearing a vintage band tee.

Prompt field:

@Sarah pushes open the door and steps inside. She scans the room, then walks
to the far end of the bar. The bartender nods.

Which descriptors belong where

Creating a Strong Soul ID Reference

The quality of your Soul ID reference image determines consistency quality.

For best results:

• Use a front-facing or 3/4 angle portrait — full face visible
• Even lighting — avoid harsh shadows obscuring features
• Neutral to slight expression — smile is fine, extreme emotion limits flexibility
• Clear image — no blur, no obstruction, no glasses if avoidable
• Solo subject — no other people in the reference frame

Image models to generate your Soul ID reference:

• Soul 2.0 — best for fashion-forward, high-aesthetic characters
• Nano Banana Pro — best for maximum photorealistic sharpness
• Seedream 4.5 — good for a range of styles

A sheet is not a menu — everything on it is a request

[FIELD — Higgsfield Studio, ZEPHYR breakdown, 2026-08-06] The intuitive way to build a sheet is to put everything on it — every weapon, every state, the mechanism drawn open — and then reference only the parts you need per shot. That is not how the model reads it. If it sees a detail, it will try to show it.

The reported failure: a mecha sheet included the cockpit hatch in its open state. The model prioritised the open hatch over the subject's standard look, so most generations came back with the hatch open when it should have been shut — or with the open and closed versions colliding and deforming the whole design. The main body being larger on the sheet did not help; presence beat proportion.

The fix is more sheets, not a richer one:

• Erase states that are not the default from the base sheet, and be especially careful with anything retractable, hinged, or deployable.
• Build a separate sheet per scenario, even when the differences look minor. Attach the extended-weapon sheet only in the shots where it is actually extended.
• Better still, give a mechanism its own close-up as a separate input when the shot has to show how it works, so it gets focus instead of competing with the full figure.

Captions on a sheet do almost nothing for video. A little label under a detail does not teach the model to operate it — you still have to describe the whole operation in the prompt. The model will not reason "retractable blade, I know how to stage that"; every mechanism gets written out as physical action in the shot that uses it.

This is the sheet-side twin of § Variety Sheets: the question is never "what could this sheet contain" but "what will the model be asked to render from it."

The Untouched Base — build a character in two passes

[FIELD — Higgsfield Studio, ONEIRIC + ADILIADA breakdowns, 2026-08-13] A character is built in two passes with two models, and the discipline is in what happens after:

1. The face pass makes the identity plate — generated in close-up, so the model captures identity at maximum detail. That close-up is the anchor every later asset of the character is checked against.
2. The looks pass builds full-figure wardrobe — costume, materials, silhouette — fitted to the already-locked face.

The two are then assembled into a sheet with one hard condition: the original close-up portrait is never run through a model again. Assembly happens in editing tools around it. Every state change — a scar, a haircut, dirt, a wound, a new wardrobe piece — is integrated point by point with masks, without touching the base.

Why it matters: the base image stays the same pixels, so identity and the skin texture that carries it survive every later version of the character. This is the structural answer to the drift documented in § Two-Tool Refinement Pipeline, where each successive model pass softens texture toward plastic — the base never takes another pass, so it never softens.

How the one-line model fix composes with it (../shared/house-rulings.md P2-3). The field fix for a flawed sheet — attach it to Nano Banana 2 with one sentence (../higgsfield-seedance-2-5/VFX-PIPELINE.md § Stage 1) — is a legal way to make the point edit. On the base, what it returns is a donor: mask the changed region back onto the untouched original, as ../higgsfield-seedance/HELL-GRIND.md also requires. Whether one full pass alone measurably softens a sheet is unmeasured here; the mask is the default because it costs minutes, while a softened base is paid for in every shot that reads it.

Hold the face, change everything else

The same rule scales to alternate versions of one character — a different era, a different universe, hero in one world and villain in the next. A new version is a new look, not a new person: wardrobe, makeup, hair, scars and damage all change, and the face stays the same set of pixels. That is what keeps an alternate version still reading as the same person, which is the whole difficulty of the job — living between "make them different" and "make them the same".

A corollary worth stating, because it is where sheets quietly rot: a new state is a new asset with a new name, never an overwrite. A character in the common room and the same character in a hospital bed are two assets of one man.

The Reference Plate — the two axes

[EMPIRICAL — Joey character-builder / banana-pro-director 3.0 skills (2026-08-16), re-derived 2026-08-22] Photoreal character work runs two things that sound like one thing and are not. Separating them is the single most useful idea in plate building, because "photorealistic" sounds like it means both and only one of them belongs on a reference.

Axis 1 — biological realism: fully ON. The subject must read as a real living person: pore texture, peach fuzz at jaw and hairline, subsurface scattering, hair strand by strand with flyaways, fabric with real weave and drape, metal with real surface, eyes with depth and moisture. This never comes off.

Axis 2 — photographic capture behaviour: OFF on every plate. A plate is not a photograph of a lit set. No key direction, no shadow side, no cast or contact shadow, no falloff on the background, no spill, no bokeh, no depth-of-field falloff, no vignette, no flare, no haze.

Why Axis 2 is off: these are references, not finished frames. Any lighting baked into a plate — a cheek triangle, a nose shadow, a contact shadow under the feet, a warm bleed on the backdrop — is inherited and amplified by every downstream generation that reads it, and it fights whatever lighting the actual scene wants. The plate carries zero lighting; the scene prompt does all the lighting later.

The trap. Writing "photographed on a real camera by a real photographer on a real set" into a plate prompt switches Axis 2 on along with Axis 1 — and Axis 2 is exactly what poisons a reference. The subject should look like a real person, rendered flat, against nothing.

The one capture phrase that survives: Photographed on a 50mm prime, even sharpness[, soft natural film grain]. Photographed not generated. — the focal length is named in plain words with no aperture, no bokeh and no falloff attached, so it buys the anti-AI-uniformity signal without switching on capture behaviour. The bracketed grain clause is an open question, not part of the settled phrase — see the box below.

The grain clause is contested — OPEN, no default (../shared/house-rulings.md P2-4). This source keeps soft natural film grain as part of the anti-uniformity signal. ../higgsfield-seedance/HELL-GRIND.md § The character sheet [OFFICIAL — Hell Grind brief] reports the opposite: bake film grain into the sheet and the character carries that look into every scene and stops reacting to new light. Both failures are inherited by every shot that reads the plate — baked grain cannot be removed per shot; a plate that reads AI-uniform, plastic skin included, softens every shot the same way (§ The Untouched Base guards the same texture) — so neither is clearly the cheaper failure, and nothing here measures it. What both sides share: Axis-1 skin detail stays fully on (Hell Grind's own sheet asks for "real skin with visible pores, no retouch"). Decide per project, pin the choice once, and never vary it across one character's plates.

The background is a FIELD, not a ROOM

The distinction most often lost. A photographed seamless is a physical surface — it takes light, it falls off, it catches spill, it holds the subject's shadow, it has a floor the subject stands on. A plate background is none of that. It is a flat uniform colour field with nothing behind the subject at all: no surface, no floor, no wall, no corner, no seam, no horizon, no plane the subject makes contact with. The subject does not stand on anything and does not stand in front of anything, and nothing the subject does affects the field.

The flattering-realism ceiling

Full skin realism is always on — but realism never means unflattering. No acne, blemishes, prominent spots, unrequested scarring, cratered pores, or aggressive detail that reads clinical. The texture is fine, soft, even, natural. Matte is the anti-plastic lever; fine-and-even is the flattering lever, and both run together. Where they conflict, resolve toward flattering — a face should always look good.

Prompt economy — lean prompts hold faces better

This runs counter to instinct. When strong references are attached, the references carry the identity load, and the prompt's job is to say what to DO with that identity in this image. Heavy visual description layered on top of a strong reference creates a double-weight prompt that dilutes the direction the model actually needs from the text.

• Identify subjects by a short distinguishing handle — "the woman with the deep red-burgundy hair" — not a paragraph re-describing bone structure the attached plate already shows.
• Put the load on what the prompt uniquely communicates: composition, framing, pose, expression, what the hands are doing, wardrobe not already in a reference.
• Drop identity description that the reference already carries, unless it is load-bearing for this specific image.
• The model reads the front of the prompt most heavily — put composition and pose there rather than burying them under description.

Rule of thumb: if a sentence re-describes something already visible in an attached reference, cut it unless the composition depends on it. The exception is the first plate — a face lock has no reference to lean on, it is the reference, and it gets the longest, most specific description of anything in this file. Everything after it gets leaner.

When two or more references are attached, say what each one carries so the model does not average them: "face, bone structure and skin tone come from the character reference; wardrobe and accessories come from the look reference."

Lean prompts next to a reference are one side of an open question for video (../shared/house-rulings.md P1-1): Higgsfield's feature-film pipeline pastes the full descriptor, word for word, beside the reference (../higgsfield-seedance/HELL-GRIND.md § The core problem). For image plates this section's rule stands; for a multi-shot video pipeline, read the ruling first.

Character Sheet Creation

A character sheet is a multi-angle reference image showing the same character from several viewpoints — typically front, 3/4, side profile, and back. It gives the model comprehensive geometry to work from and dramatically improves consistency.

Two-image floor per character [DEMO — Seedance-4K film tutorial, 2026-07]: never hand a video model a character on a single image. The minimum is one clear face view + one full body — in the tutorial's words, "so Seedance doesn't have to guess." Everything below builds upward from that floor.

Background: put sheets on neutral grey, not white or black. The law is stated once — the shade range, one pinned hex per project, the three reasons sources give for it, and the contact-shadow disagreement — in ../../templates/ad-asset-prep.md § Design for win rate. For a plate that will be read as a reference, the grey is a flat field, not a lit seamless (§ The Reference Plate — the two axes).

How to create a character sheet:

1. Generate your character in Cinema Studio using your preferred optical stack
2. Use Grid Generation (2×2 or 4×4) to produce multiple variations
3. Alternatively, use 3D Mode (Gaussian splatting) to orbit a single generation and capture front, side, and 3/4 angles
4. Arrange the best angles into a single composite reference image
5. Upload as your Soul ID reference

What to include on a character sheet:

• Front face — neutral expression, even lighting
• 3/4 angle — shows depth of facial features
• Side profile — nose, jaw, ear structure
• Full body (optional) — posture, costume, proportions
• Embedded prop sheets (optional) — bake recurring character props directly into a character-sheet panel. Higgsfield image generation handles multi-element character images well, so prop continuity often holds without a separate prop sheet. For props that don't hold this way, see § Tricky-Prop Sheets below.

Prompt pattern to generate character sheet content:

Front-facing portrait of [character description], neutral expression, even studio lighting,
clean background. Head and shoulders visible.

Then use 3D Mode to orbit and capture additional angles from the same generation.

Why it matters: A multi-angle character sheet gives Soul ID far more geometry data than a single photo. This translates directly into better consistency across extreme angle changes, action shots, and profile views.

The Hybrid Sheet — Casting a REAL Person

Everything above assumes the character is invented. Casting a real person — a client, a friend, the couple in a wedding film — is a different problem, and the sheet pipeline alone does not solve it [DEMO — Higgsfield "AI Love Stories" tutorial, 2026-08]. Feed someone's photos in as references and the sheet comes back as a lookalike actor, not the actual person: costume and body land, the face is a near-miss. It is the kind of near-miss the subject's own family spots instantly and nobody else notices, which makes it the worst kind.

Note what this is NOT. § Face-from-Wide-Shot Workaround below fixes plastic AI faces by cropping a face from a closer generated panel into a wider one — a quality repair, generated pixels throughout. This fixes identity, and the pixels come from a camera.

The hybrid, as demonstrated:

1. Generate the sheet normally. Keep it for what it is genuinely good at — costume texture, wear, silhouette, body, the panel geometry.
2. Erase the head on the full-body panels. Leave them headless. The generated face is the part that was wrong; deleting it stops it competing (the one-readable-face law: ../../templates/ad-asset-prep.md § Design for win rate).
3. Paste the person's real photograph into the portrait panel. That panel is now photographic, not generated.
4. Ride the composite as the character reference.

Why it works: the portrait panel is where a video model reads identity, and it is now a real face rather than an approximation of one. The full-body panels still carry wardrobe and proportion, which is all they were ever needed for — and a headless panel cannot contribute a wrong face, so the composite has exactly one face in it and that face is correct.

Constraints, and they are not optional. This puts a real, identifiable person into a generation pipeline. Have their permission for this specific use, keep the source photos out of anything you publish, and do not use the technique on someone who has not agreed to it. The dating, wedding and gift work this method is aimed at is exactly the context where consent is easy to get and easy to skip.

Pick the Sheet Model per JOB, Not per Project

The same sheet prompt through three image models does not produce three qualities of the same thing — it produces three different strengths, and the right answer is per-character [DEMO — Higgsfield "AI Love Stories" tutorial, 2026-08]. In that run, from one prompt across Seedream 5.0 Pro, GPT Image 2 and Nano Banana:

The AI-vs-VFX build routes "clothing, wardrobe changes, branded garments" to GPT Image 2 (../higgsfield-seedance-2-5/VFX-PIPELINE.md § Stage 1); this comparison picked Seedream 5.0 Pro for costume texture on a sheet generated from scratch. The build does not split by job — reading the two as edits on an existing sheet vs texture from scratch is a [HOUSE] inference that would let both stand — not a default. Two productions, no measurement here; OPEN (../shared/house-rulings.md P3-2), and the method below decides per character.

[UNPROVEN HERE] — one production's comparison, on two characters. Treat the method as the finding, not the table: run the same sheet prompt through two or three models and compare before locking, because the axis that decides it is whatever your character is hardest to render (hair, wardrobe wear, skin, a signature prop), and that axis changes per character. Sheets are cheap relative to the shots that depend on them; a wrong lock is paid for in every later generation.

Model specs and current pricing: ../../image-models.md.

Single-prompt 6-panel character sheet (3×2 grid)

An alternative to the multi-step assembly above is generating the entire character sheet in one prompt → one 16:9 image → 3×2 grid with six labeled panels. Same character described once; identity stays maximally locked across all six panels because the generation pass is single. Prefer this for Soul ID reference work when the image model supports a 16:9 grid layout (Nano Banana Pro and similar grid-capable models).

The six panels in canonical order:

• Panel 1 — Front body: straight-on neutral stance, full styling visible head-to-toe
• Panel 2 — 3/4 turn: body angled ~30° from camera, weight on back hip
• Panel 3 — Back body: straight back view, hair fall and accessory details from behind
• Panel 4 — Waist-up portrait: head, shoulders, upper torso
• Panel 5 — Hands detail close-up: both hands forward, ring stack, nail finish, any held prop
• Panel 6 — Face detail close-up: tight crop from collarbone up, earrings, lips, skin texture, eyes

Prompt pattern — identity described once at the opening, followed by panel position labels with what's different per panel (stance, framing, focus). Close with: "Identical character identity locked across all six panels. Uniform studio backdrop and lighting across all six panels."

Why single-prompt over multi-step: the multi-step assembly above (Grid Generation + 3D Mode + composite) produces a sheet from multiple independent generations — identity can drift panel-to-panel even with a strong reference. The single-prompt 3×2 grid keeps identity locked because all six panels render together in one pass.

Split-Panel Outfit-Change Sheet (ghost-mannequin + identity panel)

A two-panel sheet pattern for re-dressing an existing character without touching identity [DEMO — Seedance-4K film tutorial, 2026-07] — the tutorial used it to put its lead into a new Y2K outfit:

• LEFT panel — outfit only, ghost-mannequin: the full new outfit as floating garments styled as worn — invisible body, no person, no head. Describe the clothes in full here (cut, fabric, graphics, accessories, footwear); this panel owns the wardrobe.
• RIGHT panel — identity only: a large head-and-shoulders close-up from the existing character reference, declared "face matches input 100%" — same identity, features, hair, skin, no drift. This panel owns the face; wardrobe appears only as the upper edge of the LEFT panel's outfit.
• Clear vertical divide, neutral grey studio backdrop, even soft light across both panels. (The grey law — shade range, one pinned hex per project — is stated once in ../../templates/ad-asset-prep.md § Design for win rate.)

The split gives the video model one panel to read for what they wear and one for who they are, so the outfit change can't pull the face with it. Registered as a single Element, the sheet carries both. Complements § Multi- Form State Tracking below — a wardrobe state is a state like any other and gets its own sheet.

Character Anchor Block

Where § Character Sheet Creation builds the multi-angle identity reference that goes into Soul ID, the Character Anchor Block is the per-shot prompt structure that locks how that character appears IN a specific shot. The character sheet is build-time; the anchor block is shot-time.

A complete anchor block names, per character in frame, ten attributes:

1. Identity — which character this is (matches a Soul ID handle when references are present)
2. Screen position — qualitative anchor + percentage notation paired per ../higgsfield-seedance/SKILL.md § Frame Coordinate System
3. Depth layer — foreground / midground / background
4. Frame occupancy — % of frame area the character fills
5. Body orientation — direction the character faces (toward camera, away, profile-left, profile-right, three-quarter)
6. Pose — current physical configuration (standing, seated, leaning, mid-stride)
7. Gaze direction — where the character is looking, named either by frame-position or by another subject (looking at Character B)
8. Contact points — what physical surface or object the character is grounded against
9. State lock — current emotional or physical state (calm, exhausted, injured, soaked, in motion) — see § Multi-Form State Tracking for state-across-scenes discipline
10. Facial expression — specific emotional register (composed, fearful, smiling small, gritted teeth)

The block sits before the Dynamic Description in the Seedance output format and feeds a Spatial Layout Block when multiple characters share frame (see ../higgsfield-seedance/SKILL.md § Spatial Layout Block for the multi-character extension).

Multi-Form State Tracking

When a character changes state across the project — wounds from a fight scene that persist into the next scene, a costume change midway through, a transformation across multiple stages — generate a separate anchor sheet per state. Don't rely on the base character sheet plus prompt-text descriptions to track the difference; the model loses the state under iteration pressure.

The discipline is what film production calls script supervising: every shot tracks which version of the character it should match. A character with five stages (initial → fight-injuries → partially- transformed → almost-fully → completely-transformed) gets five distinct Soul ID sheets, one per stage. The shot list references the matching sheet by name; the prompt names the stage in the identity block.

The cost is one character-sheet generation pass per state; the payoff is that no shot opens a continuity bug that has to be caught at frame-review time (see ../higgsfield-seedance/FAILURE-MODES.md § Frame-level review is mandatory).