Character Asset

Create a multi-variation character with consistent identity across poses, expressions, motions, lighting, and voice — built and managed in the full-screen Character Studio — or choose an existing character from your Library or the Public Gallery.

Overview

The Character node creates a reusable character asset with a base portrait, multiple visual variation categories (expressions, poses, motions, reference angles, lighting), structured attribute and wardrobe selections, an optional high-fidelity LoRA (Cloud edition), and definition data (voice and personality). All character editing happens inside the Character Studio — a full-screen modal organized into a config-driven sidebar — while the canvas node itself stays compact and shows a summary. Characters are persisted per-project in the database and can be referenced by other nodes (scenes, image generation) via the characterRef output to maintain visual consistency.

The Canvas Node

The Character node on the canvas is a compact summary card. It shows:

• The base portrait preview
• The character name and style · gender
• A Role dropdown next to the name — the default part of this character that downstream generators take from its reference image: person (the default), face, clothes, hair, pose, expression, style, or a Custom… free-form role (e.g. earrings). See Default Role & Identity Lock below.
• An Identity lock row (Off / Soft / Strict) — how strongly downstream generators must preserve the character’s facial identity. Defaults to Soft.
• Asset-count badges for Expressions, Poses, and Motions
• A row indicating whether Voice and Personality are filled in (✓)
• A ⬡ Studio button that opens the Character Studio
• A Choose existing button (next to ⬡ Studio) that opens the Asset Picker to bind the node to a character you already have. Once a character is bound, this button becomes Replace — use it to swap in a different character.

Default Role & Identity Lock

When this character is wired into a Generate Image / Generate Video (or mentioned with @name in their prompts), its reference image is described to the model as an inline role phrase — e.g. “the person from reference image A”. The node’s two controls set the defaults for every reference derived from this character:

• Role — which part of the reference the model should take. person means the whole subject; face/clothes/hair/pose/expression/style narrow it to that aspect; a Custom… role (a short word or phrase, e.g. freckles) is used verbatim. A role picked on an individual @-mention pill always overrides the node default for that mention.
• Identity lock — Off emits no identity constraint, Soft adds a mild “preserve the overall facial likeness” line, Strict adds a strong “must match exactly” clamp. Each lock line names its reference image, so multi-character scenes stay unambiguous. A per-mention ~lock / ~nolock on the pill overrides the node default for that mention. This is the same Identity Lock field shown in the config panel and Character Studio.

Both defaults apply on image and video generation, whether the character is @-mentioned, auto-attached as an unmentioned wired reference, or added as an extra reference — including runs through the REST API, SDK, and MCP tools.

All appearance and asset editing — generating the portrait, expressions, poses, motions, reference views, lighting, and setting voice/personality — happens inside the studio. The old in-node image action buttons, version history, accordions, and asset-sheet thumbnails have been replaced by the studio.

Configuration

The config panel (right side, when a Character node is selected) is intentionally minimal:

Field	Type	Description
Summary	read-only	Character name and style at a glance.
Open Character Studio	button	Opens the full-screen Character Studio (same modal as the node’s ⬡ Studio button).
Choose from Library / Gallery	button (row)	Opens the Asset Picker to bind the node to an existing character. Becomes Replace from Library / Gallery once a character is bound — use it to swap in a different one.
Identity Lock	dropdown	Controls how strictly downstream nodes preserve the character’s face/identity (Off / Soft / Strict, default Soft). The same field as the canvas node’s Identity lock row — see Default Role & Identity Lock.
Field Mappings	section	Map upstream node outputs to the character’s inputs — {characterName} injection still works.

Character Data

In addition to the base portrait and identity settings, a character holds the following data (all built and managed in the studio):

Data	Description
sourceImageUrl	The base portrait image. The reference image for all generated assets.
expressions	Image references for facial expressions (neutral, smile, angry, surprised, sad, talking, laughing, …).
poses	Image references for body positions and stances (standing, walking, sitting, running, …).
motions	Short video clips of the character in motion, generated from the portrait via image-to-video providers.
angles	Head-angle reference views — front, 3/4 left, left profile, right profile, 3/4 right, back. (A sub-section of the Appearance page.)
bodyAngles	Full-body angles, standing naturally with arms relaxed at the sides. (A sub-section of the Appearance page.)
lightingVariations	The character under different lighting conditions — daylight / night / dramatic. (A sub-section of the Appearance page.)
person	Structured appearance attributes (hair, eyes, build, age, ethnicity, …) chosen on the Pickers page. Auto-injected into portrait + asset generation prompts server-side.
wardrobe	Structured wardrobe selections (archetype, top, bottom, outerwear, footwear, headwear, accessories, color-palette, material, era) chosen on the Pickers page. Auto-injected into portrait + asset generation prompts server-side.
voice	An ElevenLabs voice (the same voice library used by Text to Speech nodes) plus a free-form “voice traits” description. Auto-fills a connected Text to Speech node at run time (overridable per node).
personality	Four free-form text fields: Mood / Temperament, Speech Style, Movement Style, Behavioral Notes.

Voice now auto-fills a connected Text to Speech node at run time — the character’s voice, voice type, and recommended provider flow into the wired TTS node, and you can still override them on that node. Person and Wardrobe selections are injected into the character’s own portrait and asset generation prompts server-side. Personality is still stored only — auto-injection of personality into downstream generation/script nodes, and “smart” automatic selection of which expression/pose image a downstream node should use based on scene context, remain a planned follow-up. The characterRef output and Identity Lock behavior are unchanged.

Asset Categories

Category	Type	Generated via	Description
Expressions	images	image models	Facial/emotional variations (neutral, smile, angry, surprised, sad, talking, laughing, …).
Poses	images	image models	Body positions and stances (standing, walking, sitting, running, crouching, pointing, …).
Motions	video clips	image-to-video (Kling / Wan)	Short video clips of the character in motion (walking, running, waving, dancing, …). Requires a portrait first.
Reference Views (Angles)	images	image models	Head and body angles — front, 3/4 left, left profile, right profile, 3/4 right, back. Sub-section of the Appearance page.
Lighting Variations	images	image models	The character under daylight / night / dramatic lighting. Sub-section of the Appearance page.

Choosing an existing asset

Instead of building a new character in the studio, you can bind the node to a character you already have. Open the Asset Picker from either the Choose existing button on the canvas node or the Choose from Library / Gallery row in the config panel. The picker has two tabs:

• My Library — your own saved characters.
• Public Gallery — characters shared by the community. Selecting one clones it into your library first (you can’t reference another creator’s private asset), then binds the node to that fresh clone.

This works both for an empty node (first-time selection) and to replace a character that’s already set — once a character is bound, the buttons read Replace / Replace from Library / Gallery. Binding or replacing carries the full character (portrait plus every variation bucket — expressions, poses, motions, angles, lighting, voice, and personality), so downstream nodes immediately use the new character.

In two more cases the picker helps you avoid clutter:

• Already have a copy? If you pick a Public Gallery listing you’ve cloned before, the picker asks whether to use your existing copy or make a new copy — so a gallery pick never silently piles up duplicates.
• Delete from My Library. Hover a card in the My Library tab and click the trash icon to remove a saved asset. It’s archived (recoverable), and any nodes already using it keep working.

Character Studio

The Character Studio is a full-screen modal where you build and manage everything about the character. Open it from:

• The ⬡ Studio button on the canvas node, or
• The Open Character Studio button in the config panel.

The studio auto-saves as you work. Identity fields (name, description, gender, style, base outfit, voice traits, personality) are persisted ~600ms after the last keystroke. Generated assets are persisted by the backend itself at completion — every Generate call passes the character’s DB id along with the request, and the worker writes the resulting image/clip directly to the character row when the job finishes. That means if you close the tab or refresh mid-generation, the asset still lands on the character the next time you open the studio. A small “Saving… / Saved” indicator in the header reflects the current state. There is no Save button.

The studio opens on Profile and organizes everything into a config-driven sidebar with four groups:

Group	Pages
Resources	References · Pickers · LoRA (Cloud edition only)
Identity	Profile · Appearance
Visuals	Expressions · Poses · Motions · Sheet · Board
Character	Voice · Personality

Resources → References

The character’s reference photos — up to seven slots, one per framing (front face, side left, side right, 3/4 left, 3/4 right, front body, other). Real-life reference photos give the image/video providers extra conditioning so generated assets stay on-model. (Previously these lived on the Appearance tab; they’re now their own page.)

Resources → Pickers

Two structured pickers whose selections are stored on the character and auto-injected into its portrait + asset generation prompts server-side (no extra wiring):

• Person — appearance attributes (hair, eyes, build, age, ethnicity, skin, face, and more). The same Person picker available as a canvas node, embedded here and bound to the character.
• Wardrobe — archetype, top, bottom, outerwear, footwear, headwear, accessories, color palette, material, and era.

Each selection contributes a descriptive clause to the generation prompt the next time you generate the portrait or any asset, so the character renders consistently without retyping the same attributes. See Parameter Picker Catalogs for how picker clauses are assembled.

Resources → LoRA (Cloud edition only)

Train a high-fidelity character LoRA from the character’s own images, right inside the studio (relocated here from the in-node training section). The page shows a curated training-image grid: every eligible image (source portrait, reference photos, expressions, poses, head angles, body angles, lighting) is shown and selectable, all selected by default — uncheck any you don’t want to train on. A minimum of 4 images is required (the “N / 4” counter reflects your selection). From here you can Start training, watch status (untrained / queued / training / succeeded / failed), see the assigned trigger word, Re-train, or Remove the LoRA. Training is credit-metered and only available on the Cloud edition; the page is hidden on other editions.

Identity → Profile

The portrait and core identity for the character: name, portrait (with a Generate Portrait button and, for multi-candidate runs, an approve step to pick the keeper), description, gender, style, base outfit, an image-model provider picker, and a seed prompt that scaffolds portrait generation. Approving a portrait also fires an LLM caption that authors the character’s canonical visual description (used by downstream prompts). The studio opens on this page.

Identity → Appearance

Angle and lighting variations of the approved portrait, in three sub-sections, each with preset chips, a free-form prompt, an image-model picker, and Generate:

• Head Angles — front, 3/4 left, left profile, right profile, 3/4 right, back (head-and-shoulders framing).
• Body Angles — the same angle set in full-body framing, standing naturally.
• Lighting Variations — daylight / night / dramatic.

Visuals → Expressions, Poses, Motions, Sheet

• Expressions — image references. Presets: neutral, smile, angry, surprised, sad, talking, laughing, disgusted, fearful, smirk, crying. Each card has an inline-editable name (the “tag”), three regen actions (↻ regenerate same / ＋ add variation / ✏ img2img refine), and a ✕ delete. The generation bar has preset chips, a free-form custom prompt, a curated top-tier image-model picker (nano-banana-pro default, nano-banana-2, gpt-image-2, seedream), a Generate button, and a ⟳ Generate All button that queues every missing preset (it confirms first when 4+ jobs would be submitted). Every Generate button shows the current model’s credit cost inline (e.g. Generate (5 CR) / Generate All (20 CR)); preset chips show a small cost subscript so it’s clear what each tap costs.
• Poses — same layout as Expressions. Presets: standing, walking, sitting, running, crouching, pointing, fighting stance, jumping, turning.
• Motions — short video clips generated from the portrait via image-to-video providers (Kling / Wan: kling, kling-turbo, kling-3.0, wan-i2v, wan-2.7-i2v). Requires a portrait first — without one, Generate is disabled with a tooltip. Cards show video thumbnails with a ▶ overlay, an editable name, ↻ regenerate same / ＋ add variation buttons, and a ✕ delete; there is no img2img refine on motions (video refinement is a future release) and no “Generate All”. The Generate button shows the current motion provider’s credit cost inline. Presets: walking, running, waving, sitting down, fighting stance, jumping, turning around, dancing, talking gesture.
• Sheet — composite a turnaround / expression / reference sheet from the character’s assets. See the Reference Sheet node for the composition details.

Visuals → Board

Named composite reference boards — a different compositing paradigm from Sheet. An identity board composites 2–12 of the character’s own images (portrait, expressions, poses, angles, body angles, lighting, wardrobe, close-ups, reference sheets, real-life reference photos) into one dense 4K sheet using the same engine as the Image Collage node (smart layout, 4:3 target) — the strongest single-image identity anchor for downstream generations.

• New board opens a full-screen picker showing every image the character owns, grouped by category. Click to select 2–12 images; a number badge shows the collage order. Generate composites them at the Image Collage node’s 4K rate.
• Duplicate (identity boards only) reopens the picker with that board’s source images pre-selected, so you can iterate on a look without rebuilding the selection from scratch. Generating still costs the same 4K rate; the result saves under a free, auto-suffixed name (“Evening gown 2”) — the original is never overwritten.
• Rename — click a board’s caption to edit its name in place (Enter or click away saves, Escape cancels). Names that collide with another board auto-suffix, same as Duplicate. Unnamed boards show an “Add name…” affordance. Renaming applies to boards on the character’s boards list; legacy studio-app boards keep their read-only caption.
• Delete removes a board, with confirmation. It’s available for any board stored on the character’s boards list, even one without recorded source images. Duplicate always needs recorded source images. Legacy boards from the studio app — not part of the boards list — are view-only here: no Duplicate, no Delete.
• Cancel on a generating tile stops the run and refunds the reserved credits.
• Up to 12 boards per character. Like other studio assets, generation continues server-side if you close the studio mid-run — the finished board attaches to the character automatically, and its pending tile reappears on reopen until it lands.
• In reference pickers — the @ mention menu and the pill thumbnail swap picker — a character’s boards are listed first, tagged “board”.

Character → Voice, Personality

• Voice (reworked) — a full voice workspace:
  • Browse — pick an ElevenLabs voice from premade voices, the public library, and your own custom clones (with search and preview).
  • Clone from audio — upload a 30s–2min clean audio sample and create a custom voice clone, which becomes the character’s selected voice.
  • Design from text — describe a voice (“a warm, gravelly old storyteller with a slight Irish lilt”), then audition it with a preview line. Design is audition-only — saving a designed voice for reuse is a planned follow-up.
  • Selected-voice card — shows the chosen voice with a free-form voice traits textarea (e.g. “deep, calm, British accent, slight rasp”) and a preview player.
  • ▶ Talk — type a line and 🔊 Speak to hear the character speak it via text-to-speech. If the character has an approved portrait, 🎬 Speak + lip-sync portrait renders a short talking clip of the portrait saying the line. Lip-sync is credit-metered (priced per second of output — see the Lip Sync node for the rate); plain Speak uses standard text-to-speech credits.

  The selected voice auto-fills a connected Text to Speech node on the canvas at run time — the character’s voice, voice type, and recommended provider flow into the wired TTS node. You can still override them per node.

• Personality — four free-form text fields: Mood / Temperament, Speech Style, Movement Style, Behavioral Notes. Stored only — not yet auto-applied to downstream nodes.

Common patterns

• Generation bar (on every visual page): preset chips for one-tap generation of common variants, a free-form custom prompt for anything else, a model picker, and Generate. Expressions and Poses also have ⟳ Generate All. Every Generate / Generate All / preset chip shows the current model’s credit cost inline.
• Inline-editable names: every asset card’s name is editable — it’s the tag that identifies that expression/pose/motion.
• ↻ Regenerate same: re-fire the same variant on an existing card (e.g. another smile); replaces the card on completion.
• ＋ Add variation: re-fire the same variant but append the result as a new card (so you can compare versions of, say, smile side-by-side).
• ✏ img2img refine (image assets only): refine an existing expression/pose/angle/lighting image with a prompt, choosing whether to Replace the card or Add as new.
• ↑ Import: each visual page can import an existing media URL directly (no generation job) — images for image pages, video URLs for Motions.

Inputs & Outputs

Inputs:

• in – Optional text or image input for additional context (also drives {characterName} field mappings).

Outputs:

• characterRef – Character reference (identity) that can be connected to scene nodes, image generation, and other nodes that accept character references. Unchanged by the studio.
• image – The character’s portrait as a plain image. Connect this anywhere a Generate Image output can go (image References, Image-to-Image, Generate Video image input, List columns, etc.). Unlike characterRef, it carries no identity / canonical-description injection — it is just the picture.

Best Practices

• Set Pickers and reference photos before generating the portrait. Person + Wardrobe selections are injected into the portrait and asset prompts server-side, and reference photos give the providers extra conditioning — getting them in first makes the first portrait far more on-target.
• Generate the portrait first, then approve it. It’s the reference image for expressions, poses, and motions — generating assets before there’s an approved portrait gives inconsistent results, and Motions can’t be generated at all without one.
• Write a detailed description covering facial features, body type, hair, and clothing for the most consistent results (the Person/Wardrobe pickers cover the structured side of this).
• If you don’t like a generated expression or pose, use ↻ Regenerate same to get another shot at the same variant, or ＋ Add variation to keep the existing card and compare a fresh take side-by-side.
• Use ✏ img2img refine when you want to modify an existing image with a prompt (e.g. “more intense smile”) rather than re-roll from scratch — “Add as new” keeps the original while you compare.
• Curate the asset names. They’re the tags that identify each expression/pose/motion and will drive smarter automatic downstream selection in a future release.
• Use the same style setting across all characters in a project for visual coherence.
• Pick the character’s Voice in the studio, then wire the character into a Text to Speech node — the voice auto-fills at run time (override per node if needed). Use ▶ Talk to audition a line and, with an approved portrait, preview a lip-synced talking clip before committing.
• (Cloud) Once you have a good spread of on-model images, train a LoRA for the highest-fidelity reuse — curate the training grid to the cleanest shots (min 4).

Common Use Cases

• Creating consistent characters for animated explainer videos.
• Building a cast of characters for a multi-scene narrative.
• Generating character turnaround sheets (reference angles) for animation reference.
• Producing expression libraries for dialogue-driven content.
• Capturing short motion clips of a character for use as i2v references or B-roll.

Tips

• Characters are saved to the project database. They persist across sessions and can be reused in multiple workflows within the same project.
• The characterRef output carries the character’s identity information, allowing downstream nodes to maintain visual consistency.
• Each expression, pose, angle, and lighting variation is stored as an individual image, so you can reference specific assets.
• Pickers (Person + Wardrobe) and reference photos affect only the character’s own portrait and asset generation — they’re applied automatically server-side, so you don’t wire them anywhere.
• Motions are short video clips — a motion clip costs the same as the equivalent image-to-video generation on the chosen provider (e.g. a Kling clip costs the same as a Kling image-to-video). See the Image to Video node for provider details.
• Use ↑ Import to bring in artwork or footage you already have instead of generating it.
• Close the studio to collapse it back to the compact canvas node — your saved data stays on the node and in the project database.