Visual styles

A look your show uses -- the words it adds to prompts and the models it recommends -- so a scene can look different from the rest of an episode without setting anything on every shot.

A visual style is a look your show uses -- "Anime", "Claymation", "Dream". It holds two things: the prompt trigger text that look adds, and the models it recommends for the operations where that look matters. A show normally has one or two. Give a scene a style and every shot in it runs that way.

A dream sequence can be soft and bleached while the rest of the episode is sharp, and you set it once, on the scene -- not on each of its shots.

You do not need a style. A scene or shot can have none. With no style OA adds no look of its own to the prompt -- you get the model's own look -- and when you perform an operation OA lists every model that can do it and you pick one. Nothing is blocked for want of a style.

A style recommends a model for an operation. The recommendation is a workflow (a model at a build) plus, optionally, prompt trigger text. When you perform that operation on a shot in the style, the recommended model is already selected.

You do not set a style up per operation. A recommendation covers everything its workflow does. Most workflows do one job; some do several -- an image-edit workflow can remove a background, change a pose and relight a shot, all steered by the prompt. Each card lists everything it covers -- Covers: Image to Image, Pose Change, Remove Background -- so nothing changes behind your back.

Several recommendations for one operation, and one default

A style can recommend more than one workflow for the same operation -- two ways of making an image, say. One of them is marked with a green check: it is that operation's default, and it is the one preselected. The others are alternatives: offered wherever you can change the choice, never used unless you pick one.

This is not an error, and nothing is shown in red. When you add a workflow for an operation the style already answers, the form says so plainly -- Anime also has Text to Image (Flux Klein 2.0 4B) for Text to Image; the one marked default is used -- and saves it.

The default is per operation, because a recommendation covers several. An image-edit workflow can be the default for four of the operations it performs and an alternative for the other two; when that happens the card says which -- Default for: Image to Image, Pose Change.

The rules are short:

  • The first workflow to cover an operation becomes its default automatically.
  • Adding another for that operation leaves the default where it is.
  • Choose Make this the default from a card's menu (the dots in its top-right corner) to move it. That is the only thing that moves the default, so there is always exactly one answer.
  • Removing the default makes the oldest remaining workflow that covers the operation the default. Only when nothing else covers it does the operation have no recommendation -- and the confirmation says which of those will happen before you press Remove.

The prompt trigger text

Prompt trigger text is a piece of a prompt, never a whole one. It carries what the look needs said -- a style LoRA's trigger word, "flat cel shading, bold outlines" -- and the assistant folds it into the prompt it writes for each shot, along with the shot's own content.

Leave it empty and the recommendation is just "run this workflow". Negative prompts do not go here: they are baked into the workflow itself.

A visual style says what the picture LOOKS like. Who is IN it is a separate question, answered by the cast's identity LoRAs -- a character, location or prop trained into a small add-on file and uploaded into that look's gallery. The two meet at the run: the style recommends the workflow, and the cast's in-use identity LoRAs fill whatever slots that workflow has. Whether a workflow has any is a fact about the workflow, so the same shot can carry a likeness under one style and say "This workflow takes no identity LoRAs" under another.

What a style says about the look

A style's name is what you call it ("Anime"); its description is the words for what it looks like ("flat cel shading, clean ink outlines, bright saturated colour"). When OA writes a prompt for you -- Write prompt, or a prompt generator -- it writes the look from that description, in the way the chosen model likes its prompts written. It never borrows a look from the project's genre or tone: those describe the story, not the picture.

With no style, or a style with no description, a prompt leaves the look to the model's own default.

Visual style notes, and who defines the styles

The project's Visual style notes are the intent, written in The Foundation: "2D anime for most scenes, claymation for comic shots". They guide the AI when it chooses among the styles you have defined. The AI never creates a style on its own.

Styles are defined at Choose Models, in The Visuals, after the script and before The Boards -- one for each look the notes call for. The assistant can suggest which styles to create from the notes. You can also create a style at any time on the Visual styles page: a Dream style added mid-production, say.

Where a style applies

Two places, in this order:

  1. A shot's visual style, if it has one.
  2. Otherwise its scene's.

If neither has one, there is no style: nothing is added to the prompt and you pick the model. There is no project-wide default style.

On a shot, the field reads Inherit from scene (Dream) until you set one, so you can always see what it is getting; if the scene has none it is blank. Setting a shot to a different style is an override. That is how one shot inside a dream sequence goes back to the episode's ordinary look. You can clear a style any time.

A gallery that is not a shot's -- a character sheet, a location plate, a prop -- has no scene to take a look from, so it has no style unless you choose one. The operation panel there offers a style switch, for when you want to render one reference in the dream look.

Scene Planning sets them

Shots are created at The Boards, in Scene Planning. That step is best done after Choose Models (recommended, never required). When the assistant plans a scene's shots (Plan shots or Replan shots) it also proposes the scene's style, and a style for any shot that differs. It chooses only from the styles you have defined, using the Visual style notes and the screenplay, and gives a one-line reason for each. A scene with no style simply gets the proposed one; a scene that already has a style is never overwritten without asking, and a shot's style is always a proposal, even when it was blank -- the proposal waits (it survives a reload) and shows the style's name, so you can keep or change it. If no styles are defined yet it leaves them blank and says so. You can change or clear any of it.

What a run tells you

Before you press Queue, the run row says where the loaded workflow came from. When it is the one your look recommends it names the look -- Dream, from Scene 4 -- and that is the default, already selected for the operation you chose. When you have picked something else it says Chosen for this run instead, and names no look: the look did not choose it.

You can run something else. The operation picker lists every available workflow, and picking one there changes this run only -- the style is left exactly as it was, and the next time you open that gallery it is back on the workflow your look recommends. Nothing is labelled "adjusted", because there is nothing to label: the tile records which workflow actually ran.

The run record is the history, not the style. A style is a living definition you can change any time; Run again reproduces what the record says, not what the style happens to say today.

Recommend for a style, after you pick

When the shot has a style and you pick a model it does not recommend for the operation, the pick offers a tick: Recommend for Anime. Tick it and the model is saved into that style, so it is preselected next time. Leave it unticked and the pick is for this run only. When the shot has no style there is nothing to tick.

Trying a workflow your look does not recommend is not a decision, and the app does not treat it as one. You run it, you look at the result, and only then -- if it is the one you wanted -- do you say so. Picking one of those opens a reminder rather than a question: Try Flux Klein 2.0 4B, and under it Flux Klein 2.0 4B isn't approved for Anime yet. Try it here -- if you like the result, approve it from the image's menu. Press Use and it runs. Nothing is written.

Then, on the image it made, the menu carries Approve for Anime... -- named for the look this shot is in. It takes the workflow that actually made that tile, not what the look recommends now and not what the panel happens to be showing. The dialog shows the prompt the run used so you can lift the piece of it the look needs into the trigger text; it does not copy the whole prompt in, because a whole prompt would ride along on every future run.

Adding adds the workflow to the look. It becomes the default only for operations the look has no default for; where the look already has one, that one stays and the dialog says so. The default never moves because of something you did somewhere else. Below it, Add to another visual style... is the same thing for a different look, or for an image that stands in no shot at all; it asks which.

The Visual styles page

Your project's Visual styles page lists your styles down the left; select one to see its recommended models on the right. Each card says the workflow, the operations it covers, and its prompt trigger text. A green check in the top-left corner marks the default for the operations it covers; the card's menu, top-right, holds Make this the default and Edit.

  • + New visual style names a look. It starts with no recommendations -- every operation lets you pick from all models -- and you add the models that make it different. Every style can be renamed and deleted.
  • Approve a workflow picks an operation, then a model that can do it, then the workflow (its build, variant, or your own copy). The operation is how you find the workflow; before you save, the form says which other operations it will also cover, and whether this style already has something for one of them.
  • From a model style... starts from a look instead of a graph. See below.
  • Choose Make this the default from a card's menu to make that workflow the default for the operations it covers.
  • Press Edit in a card's bottom-right corner to change its workflow or its Prompt trigger text, or to Remove from <Style>. Remove asks you to confirm, and the confirmation says exactly what will change: another workflow becoming the default, the operation having no recommendation, or nothing at all.
  • Hover a style's name for its pen (rename) and trashcan (delete). Any scene or shot set to a deleted style becomes blank; the runs it produced keep their own record.
  • The description is for you -- "soft, bleached, slightly out of time" is the kind of note that is worth writing. Hover Description for its pen, type, and press Save.

This page is the one place to review and edit what a style recommends. The AI Models page shows it read-only -- Recommended by: Anime (image to video) on a model -- with a link here.

If you have used the app before, the choices you already made are kept: an old Standard style that holds models now shows as an ordinary style called "Standard", which you can rename or delete. One that holds nothing is not shown, and scenes that had no style are simply blank. Nothing was moved and nothing needs tidying up.

From a model style

A model style is a look we built for one model -- Watercolour on Flux Klein 2.0 4B, say -- with example images and the prompt trigger text that produces it. A model style answers what can this model look like?; a visual style answers what does this project look like here? The model styles live in each model's card, in the AI Models catalogue, where you can see the pictures before you commit to anything.

There are two doors into the same thing, and both end up adding one recommendation:

  • Approve for a visual style... in a style card's menu, over in AI Models. Best when you are browsing looks.
  • From a model style... beside Approve a workflow on this page. Best when you already know the look you want and are standing in the visual style that needs it.

Either way you pick the build and the look's prompt trigger text is filled in for you, editable before you save. If the model style has no workflow at the build you picked, the form says so and lists the builds it has.

This follows the same default rule as everything else here: it becomes the default only for operations this style has no default for, and where the style keeps its existing default the form says which, and where to change it.

You do not have to open the Visual styles page to set one up. In any gallery's Perform operation panel, press Choose an operation:

  • If the shot's look recommends a model for the operation, its card has the green check and says Dream's default. Pressing it runs. It is first in the list, and it is the one the run row had already selected.
  • Other workflows the same look recommends for that operation come next, each saying Approved for Dream. They are offered, never chosen for you.
  • Then every other available workflow.
  • With no style, or an operation the style has nothing for, nothing is preselected: every model for the operation is listed and you pick. Pressing a model asks which build, and runs. If the shot has a style you can tick Recommend for <Style> there to save it.
  • Other builds on a row is the offer to run the same model at a different build.

Picking anything else here is for this run only unless you tick Recommend. Close the gallery and reopen it and you are back on the workflow your look recommends. Changing what a look recommends is the Visual styles page, the tick, or the image's Approve for... menu. Nothing else.

A workflow in your Workflow Library carries Use for visual style... in its menu too: pick a look, and that graph is added to it.

Deciding an experiment

An experiment ends with you saying which result won. Decide asks which look that answer belongs to and adds the workflow that made the winning tile -- as that look's default for the operations it covers. What the look recommended before stays as an alternative. An experiment run to settle one scene's look should not quietly change every scene's, which is why it asks. If you have defined no styles, only your note is recorded; create a style first on the Visual styles page to give the winner somewhere to go.

So does the Assistant

The Assistant works the same way. Ask it for a keyframe on a shot and it sets up keyframe -- the shot's look decides which workflow runs it, and adds that look's prompt trigger text and the cast's likeness to the prompt. You are not asked which model when the look has a recommendation, and it does not choose one for you: it is the same answer the run row would have shown you.

Name a workflow yourself and that wins, for that run, exactly as picking one in the gallery does. And where the look recommends nothing for the operation, the Assistant says so and asks you to pick -- it will not pick one to get the job done. It never creates a style unasked.