Asset library: characters, locations, props

The media library answers "what is this file"; the asset library answers "who is this". An asset is a named thing — a character, a location, a prop — whose reference images are images from your media library: the asset only points at them, it never copies the files. The same character usually appears across projects, so assets belong to the workspace, not to a project; tags keep them organised.

Open Assets next to Media in the sidebar. Tabs at the top switch between characters, locations and props; each card is an asset's cover.

Organise your assets#

The asset library uses the same toolbar as the media library, and it sticks to the top as you scroll:

• Search names, descriptions and prompt descriptors; sort by recently updated, name, or most references first; filter by tags, choosing "has all of them" or "has any of them" when you tick several.
• Right-click a card to open, rename, edit tags, add a variant or delete it.
• Select switches to multi-select: clicking a card selects it instead of opening it, and you can select all. Then add tags to every selected asset (existing tags stay) or delete them (variants go too; the confirmation says so).

Create an asset#

Click New character (or location, prop), give it a name, and you land on its detail page. The top shows who it is: a large cover, then the name, kind, a one-line description, tags, and a quick summary (how many references and variants, voice, blockout color, real or fictional). The name, description and tags read like text; click to edit, and they save when you leave them. A character's header also has Make it talk: write a script and it is voiced with the character's own voice and turned into a talking video from its front image (see Digital humans).

Below that, tabs split the rest:

• References: the asset's reference wall, see the next section.
• Variants: its winter clothes, teenage version…, see "Variants".
• Settings: the prompt descriptor (for models — for a character: face, hair, clothing, age; for a location: layout, furnishings, light, time and weather; for a prop: shape, color, material, wear. Whenever the asset is @-mentioned in a generation, this text is appended to the prompt) and the kind-specific fields — characters have a voice (from any voice engine — your cloned voice library, Edge or a cloud engine; pick the engine, then the voice; with Alibaba Bailian (CosyVoice) the list also has your declared library voices as "Cloned · name", spoken from their copy in your Bailian account — see Editing → Your own voice through Bailian), a blockout color for their stand-in figure in 3D blockouts, and real or fictional with a consent declaration; locations have a linked 3D scene and "time / weather / light"; props have their 3D model.
• Where it's used: see the last section.

The reference wall#

Each reference image is labelled with its angle or purpose, and each kind has its own set:

Kind	Labels
Character	front, side, back, turnaround, close-up, full body, expression, concept, detail
Location	wide shot, reverse angle, overhead, concept, detail — a location's "angles" are camera set-ups, with no front or side
Prop	front, side, back, turnaround, close-up, concept, detail

• Add from media library opens a picker: search by name, filter images or videos, click to select several and attach them at once — in the order you picked them. Images already attached are listed too, marked Attached.
• Upload, or drop files on the wall: they are imported into the media library first, then attached.
• Change the angle by clicking the label on the image's top-left corner. Set as cover, move earlier / later and remove from references are in its ⋯ menu; you can also drag images to reorder.
• Click an image to enlarge it; the lightbox pages through the whole wall. The cover at the top of the page enlarges too.
• Without a cover, the card uses the first reference image.

AI fill-in#

AI fill-in at the top right of the reference wall draws more images of the same asset from its existing references and attaches them back by angle:

• Fill in angles: for characters and props, the front, side, back and turnaround views it doesn't have yet; for locations, the wide shot, reverse angle and overhead view. You can choose to redraw every angle instead.
• Generate expressions (characters only): list a few expressions (comma-separated, up to 8; empty draws smiling, laughing, surprised, angry and sad) — one image each, attached as Expression.

Pick an image model that takes reference images (text-only models can't draw the same one, so they aren't listed); leave it empty to use your default image model. Each image is one image generation, billed by the model; the button reads "Drawing…" while it runs and the task center tells you when it's done. An asset with no image reference needs one first.

Deleting an asset never deletes media: the images stay in the media library. Deleting media removes that reference from any asset that used it; the asset itself stays, and its detail page says which image went missing until you dismiss the note. The delete confirmation in the media library mentions this too.

The media detail view gains a Used by assets row, and right-clicking an image in the media library offers Set as an asset's reference… (optionally as its cover).

Variants#

Winter clothes or the teenage version of a character, the same place by day, on a rainy night or years later, the same object brand new, worn or in another color — these are variants hanging under it. A variant inherits its parent's prompt descriptor (characters also inherit the voice); only the references and this version's extra description change — "Zhang San · Winter" is the same person as "Zhang San". A variant without references of its own uses its parent's in generations. Deleting an asset that has variants asks you to confirm that they go too.

Real people and consent#

A character is declared fictional or a real person. A real person needs "This is me" or "The person has given consent" — every option explains what it means on screen. The server records who declared it and when. A real person without a declaration can't be used for digital humans ("make it talk", or its reference images as a talking photo); this is the separate consent that deep-synthesis regulations require.

@-mention assets in generations#

When you write a prompt in a board's generation item, the @ menu lists characters, locations and props from the asset library (with cover and kind) alongside media; AI Studio has an @ Asset button under its prompt; the workflow AI generate node has an Assets field; the agent's generate_image / generate_video take entity_ids. All four do the same thing:

1. The asset's prompt descriptor is appended to the prompt (a variant contributes its parent's descriptor plus its own).
2. Reference images are picked in order: for characters turnaround > front > full body, for props turnaround > front, for locations wide shot > concept, then the rest in wall order. With several assets mentioned, picks take turns so the first one doesn't use up every slot.
3. Only the selected model's description decides how many fit: how many reference images it takes and whether it can combine them with an attached first/last frame. A model without a known description gets none — nothing is guessed.
4. What didn't fit is reported: the AI Studio history and the board both say how many images were attached, how many were left out, and why.

Models that build a reusable subject first (such as Kling) get one subject per mentioned character, built from its reference images with the front view first. The subject's name is derived from the image contents, so using the same character again finds the same subject instead of creating a duplicate.

Assets in workflows#

Workflows have three asset-library nodes under the "Assets" group of Add node:

• Get from asset library: pick an asset, or look it up by kind + name. It outputs the prompt descriptor, references (in the order generation picks them), cover and voice, ready to feed generation, plugin and other nodes. When a name lookup finds nothing, Found is 0 — branch on it with a Condition to draw it first.

• Save to asset library: save a set of drawn images (a turnaround, concept art…) as a character / location / prop. If one of that kind and name exists, it's merged: only missing references are added, and the prompt descriptor and description are filled only if empty — edits by people aren't overwritten. You can also choose "always create a new one".

• List library assets: lists one kind of asset as one-per-line text (name — prompt descriptor), usually handed to an LLM so it can tell whether someone in the story is already in the library.

Under "AI" there are also Fill in angles and Generate expressions — the same thing as AI fill-in on the reference wall.

Together they "reuse what exists, save what's new". The Topic to finished video template works this way: before it defines the cast and locations it shows the model what the library already has, and reuses a name exactly when the story needs that one; each character and location is looked up first — if it has images they're used as-is, and only new ones get a turnaround or concept art, which is then saved to the library. The same character in your second film has the same face.

The asset item on boards#

Add → From your libraries → Asset places an asset item on a board: cover, name and a kind badge.

• Connect it into an image or video generation item and it works exactly like an @ mention; the generation panel shows the connected assets along its top, and they come first in the @ menu.
• Connect it into a note or a document and "Let AI write" reads its description and looks at its first two references ("describe this character", "write her backstory").
• Select the asset item and its toolbar offers its abilities: Fill in angles (all three kinds), Generate expressions (characters only) and Make it talk (characters only; see Digital humans). The new images appear to the right and are attached to the asset.

Where it's used#

The Where it's used tab lists where the asset is used: boards that hold its asset item or mention it in a prompt, generations that named it, and workflows whose generation nodes use it. Click a row to go there.