Reference sets: the source of truth for your art pipeline

Last updated: October 11, 2026

A reference set is where your studio keeps a concept it reuses: a character, a house style, a prop, a location, a brand, a motion. You curate it once, and every generation, edit, workflow and teammate that uses it draws on the same files, wording and colors. Training is one thing you can do with a set, not what a set is for; most of the value comes from having one agreed source of truth that the whole pipeline points at.

The clip shows a set being made from three images. For the steps, see Create a reference set.

What a set holds

Open any set from Reference Sets in the sidebar. Its editor has a tab for each part of the concept:

Reference files

The images, and for the models that use them video, audio or 3D files, that show the concept. The first image is the set's portrait: some models take only one image per set, and that is the one they get.

Category

What kind of concept the set is (see below). It decides which models the set means something to.

Name and description

The name is how you call the set in a prompt (@ followed by the name). The description is written text Layer can fall back on when a model cannot use anything else in the set.

Color palette

Named colors, each with an optional role. They are added to the prompt of every generation that uses the set.

Prompts

Always applied text joins every prompt that uses the set, so nobody has to retype the house phrasing. Prompt templates are saved snippets that people insert on purpose, with Insert prompt template next to the prompt box; they are never added on their own.

Skills

Instructions offered to the agent whenever the set is used.

Brand

On a Brand set, or any set holding them: Logo mark, Wordmark and Fonts. These never steer a generation on their own; the agent reads them and places a logo only when asked.

Finetunes and Runs

Models trained from the set, if any, and the history of every training run.

The kinds of concept a set can hold

You pick the category with the chip above the set's name. Each one tells Layer which kinds of output the set is relevant to. A set attached to a model it does not apply to is hidden from that model's form and left out of the run, so a character set never leaks into a sound effect.

Layer's own library, under the Featured scope on the Reference Sets page, adds sets Layer maintains for every workspace: art styles, characters, animations for 3D, and a library of film techniques for video, grouped in video forms as Camera, Staging, Look and Performance. The film technique categories are Layer's own; you cannot file your sets under them.

Art Style

A look. Rendering, line, lighting, finish. Relevant to every kind of output. New sets start here.

Character

A hero, an NPC, a mascot, kept on-model across poses and angles. Used for images, video, 3D, audio (a voice) and playables.

Object

A prop, a weapon, a vehicle, an item icon family. Used for images, video, 3D and playables.

Scene

A location or environment. Used for images, video, 3D and playables.

Animation

A motion a rigged 3D mesh plays. Used by 3D models only, one per generation.

Brand

Logos, fonts and a palette. Used for images, video, text and playables.

How Layer uses a set on a run

Models accept very different inputs. One takes trained weights, another takes several reference images, another reads text only. You attach the same set to any of them, and Layer translates it into whatever that model can take, in this order:

  1. A finetune trained for this model. If the set has an active finetune for the model you picked, the run applies it, and a weight control lets you dial its strength. A finetune belongs to the model it was trained on; on any other model, Layer moves to the next step.
  2. The reference files. Layer places the set's files into the model's own reference inputs (style references, character elements, reference audio) as far as the model has room. For a video model that takes a first frame, a Character set with nowhere else to go opens the shot on its portrait, unless you gave a first frame yourself.
  3. The description, as text. When nothing else applies, the set's description replaces its @ mention in the prompt, or is added to the end.

On top of whichever step applied, the Always applied text and a color palette line are added to the prompt. Audio models skip the palette. Models that speak the prompt aloud (text to speech, talking avatars) are the exception: the prompt is the script, so no set text is added and set mentions are removed from it.

Because the result depends on the model, the form tells you how well your choice fits. When you attach a set, Layer can switch to a model that applies it more directly and says so, for example "applies this reference set as a trained LoRA" or "using your reference images", with a button to switch back. If your model can only approximate the set through the prompt, or drops it, a warning names a model that matches it directly. On the Reference Sets page, the Compatible with filter shows which sets a given model can use. For more on picking, see Choose a model.

Where sets plug in

Tool forms

The Reference set button above the prompt attaches sets. Some models add slots of their own beside it, such as LoRA for trained sets, Animations on 3D models, or the film technique slots on video models.

Prompts

Type @ in a prompt to mention a set. Suggestions favor the sets in your current project that the chosen model can use. The mention also marks where in the sentence the set's text lands.

From the set itself

Use in session in the editor's header opens a new session with the set attached and a model that works with it. Its menu lists the set's finished finetunes and its other kinds of output.

The agent

Mention a set in chat and the agent works with it, including its skills. While you have a set open in the editor, the chat is scoped to that set. The agent can also create and update sets, and start training. See Write a brief for the agent.

Workflows

Generation nodes take reference sets, and a Reference Sets node holds a list of sets you can connect to several generators at once. Each set in a node's slot runs as its own generation, so one workflow can produce the same shot for every character in a cast. See Find and run a workflow.

Blender

The add-on picks an Animation set when it animates a rigged mesh.

MCP and the API

Agents and scripts can list, read and create sets, pass them to a generation, and start training, so automated pipelines use the same sets as your artists. See Developer tools.

Where training fits

Training turns a set's files into a model trained on them: an image finetune, an edit finetune learned from before-and-after pairs, a custom image style, a voice, or a 3D motion, depending on what the set holds and which models can train. It is worth it when you need the tightest consistency on a model you use heavily. It is not required: an untrained set already works through its files, text and palette on every model that takes them.

  • Training costs Creative Units, and needs a role that can train (below).
  • Each finetune is tied to the model it trained on. A set can hold finetunes for several models, and uses each where it applies.
  • Retrain makes a new version and switches to it once it is ready; the current one keeps working until then. When several versions exist, Use this one picks which the set uses, so rolling back is one click.
  • When images join the set after an image finetune trained, its row in Finetunes says how many are new and offers Retrain.
  • Weights trained elsewhere can come in with Upload finetune, and finished weights can be downloaded by roles allowed to.
  • The set's strength meter climbs as you add references, but it does not reach Strong without a trained finetune.

Sharing and control in a workspace

Sets belong to the workspace, not to the person who made them, so everyone draws on the same version. On the Reference Sets page, the scope filter separates the Workspace's own sets from Featured ones, and Type, labels and projects narrow the list. Add sets to a project so the team on that game finds them, and suggestions in that project rank them first.

Roles decide who can change the source of truth:

Custom roles use the same permissions: Edit reference sets, Train styles and Download style weights. See Invite teammates and set their roles.

Admin, Editor and Art Director

Can create and edit sets, train them and download trained weights.

Artist

Can create and edit sets but cannot train.

Member

Cannot edit sets. Anyone who can view a set but not edit it sees it read-only.

Building a library for a live-service game

  • Start with what gets reused. The house style, the main cast, recurring props, the core locations, the brand. Each becomes a set before production volume starts.
  • Combine sets, do not merge them. Attach a style set and a character set to the same run and each does its own job. One set per concept means a new season can swap the location set while the style and cast stay put.
  • Organize by game and season. Use labels and projects so the right sets surface in the right project, and new artists find them without asking.
  • Put the sets into workflows. A workflow that already names the style and cast produces on-model output whoever runs it.
  • Update in place. When art direction moves, add the approved pieces to the existing set and retrain. Everything that points at the set picks up the change, and Runs keeps every earlier version.

Designing a good set

  • One concept per set. A style set shares a look across varied subjects. A character set shows one character across poses, angles and expressions. A set that mixes both teaches neither.
  • Include only what you would ship. One off-model image pulls the set off course. Crop out UI, watermarks, signatures and placeholder text unless they are part of the concept. Use the highest resolution you have.
  • Order the files. Put the clearest, most representative image first: it is the portrait a model gets when it takes one image per set.
  • Name it for the prompt. People type the name after @, so make it short and unambiguous ("Captain Rhea", "Frostpeak village"). Renaming later is safe: a finetune keeps the name it was trained with.
  • Write the description as prompt text. It is what models that cannot read the files get instead, so describe the look or subject in the words you would put in a prompt.
  • Set the category. A set left as Art Style is offered everywhere; the right category keeps it out of runs where it means nothing, and earns its own slot on the models that have one.
  • Pin the fixed parts. Put exact colors in Color palette and non-negotiable phrasing in Always applied, so they hold even on models that cannot read the files.
  • Branch with Duplicate. For a variant (a seasonal skin, an alternate art direction), duplicate the set from the ⋯ menu instead of editing the original.

Common misconceptions

"A reference set is a training dataset"

Training is one use. Files, description, palette and always applied text all work on runs without any training.

"A trained set works the same on every model"

A finetune applies only on the model it was trained for. Elsewhere the set falls back to its files and text, and the form warns you when only the prompt is left.

"More images always help"

Consistency matters more than count. The strength meter flattens as files are added, and an off-model image does more harm than a missing one.

"Attaching a set is the same as attaching its images"

A set brings its text, palette, finetunes and skills too, and is translated for each model. Loose images attached to one generation carry none of that and are not shared with the team.

"Every attached set is used"

A set whose category means nothing for the model is left out, and a model can run out of room for reference files. Check the form's notices before a large batch.

Related