Creative boards

A creative board is where you lay ideas out and work on them in plain sight: put media, text and generations on one desk, then look, compare and revise until an idea takes shape. The content is the interface, and placement and groups are how you organize it; try one prompt on several models side by side and carry on with the one you like; every item remembers how it was made, so you can regenerate or run it again. When you want to lock a process in and have it run again by itself on new input, use a workflow.

A board holds nine kinds of item: note, document, image, video, audio, 3D scene, asset, timeline, and the group that fences others in. Connecting two items passes upstream content to a downstream input — a reference image to a video item, say, or a document to a note — a line means "refer to this", not "then do this". Drop a connection onto the visible + area on either side of an item. Select an item to open its action bar above and its generation or editing panel below.

Getting things onto the board#

Add on the toolbar (⌘N) is a searchable list — each row a name and a one-line hint — in three groups:

Group	Items
New	Image, video, audio, note, document, 3D scene (connect a script to build a blockout), timeline
From your libraries	Media (an image, video, audio clip or document from the library), asset (a character, location or prop from the asset library), 3D scene (reference an existing scene)
Organize	Group

• Drag from Finder or Explorer: images, videos, audio and PDF / Word / slide files go into the library first, then land as an item of their own kind where you let go.
• ⌘V / Ctrl+V: board items (see below), screenshots and copied media files become one item each, and plain text becomes a note in the middle of the view. While the cursor is in a note, a prompt box or any other text field, pasting goes there as usual and nothing is added to the canvas.
• The media picker is a grid of thumbnails, newest first; under each one are its name, kind and size, and where it came from and how long ago (so several assets with the same name tell apart). It searches by name, filters by kind, can show only what's On this board, and imports directly with Upload a file or by dropping files onto it (Drop to upload) — it only takes the kind this item needs. Click a tile to place it, or use the keyboard: Enter in the search box picks the highlighted tile, ↓ moves into the grid, the arrow keys move between tiles and Enter picks. Image and video tiles have a full-size view in their top-left corner.

What a connection gives you#

Connections are not decoration. The downstream item reshapes its panel around what it receives:

Upstream → downstream	What arrives
Note → image / video	The text fills in the prompt
Note → audio	The text is what gets read aloud
Note / document → note	The upstream text is sent as material, kept separate from your instruction
Document → image / video	The backend reads the document's text to the model along the connection; one still parsing or failed to parse is reported
Image → image	Attached as a reference image
Image → video	One image becomes the first frame; several become references. When the selected model doesn't take images, the panel says the connected image won't be used
Image / video → note	The model looks at it — images directly, video as sampled frames
Audio → note	Its transcript, if one exists, becomes material
Asset → image / video	The same as @-ing it in the prompt: its description joins the prompt and its references attach, as many as the model takes
3D scene → image / video	A 3D reference chip appears on the panel; see "3D scene items" below
Video / image / audio → timeline	Appended to the end of that timeline

What flows in never overwrites what you typed: it only fills a box that is still empty, or one that still holds the previous auto-filled text.

Connecting several at once. Select two or more items and the selection frame grows a + with a count on its right edge. Drag from it: release on an item and every selected item connects to it (each line previews whether it can connect), or release on empty space for "Add a cell linked to the N selected", a new item that takes all of them as upstream — that's how several images become references together. While several are selected, each item's own + only shows on hover and connects just that item. Either way it's one undo step.

Copy and paste items. ⌘C copies the selected items and the connections between them, ⌘V pastes; connections to items you didn't copy are left out. Running state isn't copied, and a timeline item pastes as a separate copy of its timeline, marked as a copy.

Top toolbar#

From left to right: Add (⌘N) with undo / redo; comment mode, show / hide comments, the discussion center; marker mode, show / hide markers, the marker list; Find node (⌘F), edge style (Bezier / Smooth step, remembered per board), trackpad / mouse mode, overview and fit to view; then the AI assistant and More (delete the board, with confirmation) at the right end. In narrow windows, arrows scroll through the tools; the AI assistant and More stay visible.

⌘F finds items by text, name or type (named items are listed by their names): matches are outlined on the canvas, Enter / ⇧Enter jumps to the next / previous one, Escape closes. Rename the board by clicking its name in the path at the top.

In the board list, a card's ⋯ button and its right-click menu offer the same actions: open, rename, duplicate and delete (with confirmation). Select in the header enters multi-select with select-all and bulk delete; if some deletions fail, the message says how many and those cards stay selected. A duplicate leaves out items that are still generating: they become empty slots with their prompts kept.

What each item can do#

Select an item: the action bar sits above it, its panel below. Every panel looks the same: a row of upstream chips and attached media on top, the prompt or the text to work on as the body, and a bottom bar with a few setting chips (model, voice, target language…), a Settings button holding the rest, and a round send button (⌘↵) on the right. However long the body gets, the bottom bar stays pinned.

• Image / video — while empty, a generation panel hangs underneath. Each row of the model dropdown reads "model · connection"; your default model from Settings is preselected, and with none set it says Choose a model; switching models resets the settings to the new model's defaults. Aspect ratio, resolution, duration, the generation mode (first & last frame / all-round reference) and sound are driven by the model's description and live under Settings. When a model can make several at once, the bottom bar gains Results per run (N×), and that many placeholders are laid out to the right before it runs. Image items also offer Prompt templates. Once there is a result, the action bar has Preview, Trim and Swap, then the item's abilities (next section).
• Reference media — the + slot opens Add reference material, saying how many more each role takes; you can also drop files onto the slot or press ⌘V in the panel to paste a screenshot, and it's uploaded and attached.
• Note — selecting a note does not open the writing panel; you usually select a note to move it or recolor it. Press Ask AI to write in the action bar when you want help: empty means "write it", with text it means "change it the way I say"; press it again to close. Double-click still edits by hand. The bar also has colors and Save to notes. A note placed with Add to board from a note's selection toolbar shows "From note “…”" underneath; clicking it goes back to that note and finds the passage (see Notes).
• Audio — pick a voice and the text to read (voice-over). With Alibaba Bailian (CosyVoice) you can also pick from My cloned voices, spoken from the library voice's copy in your Bailian account; the first time, you're asked whether its reference clip may be uploaded (see Editing → Your own voice through Bailian). The action bar above an empty audio item switches between ways to fill it: Voice-over, Generate (music, sound effects) once a generation model supports audio, and generators from plugins (extra ones go under "…").
• Group frame — fence things off and name them; with Move together on, dragging the frame carries what's inside. Select a few items and press ⌘G to group them directly.

"+ Make ▾" for the next step. On an item with content, + Make ▾ spells out what comes next: a note offers "Turn this text into an image / a video" and "Read this text aloud"; an image "Use this image as the first frame of a video"; a video "Make a video that references this one"; audio "Make a video to go with this audio"; a 3D scene "Generate an image from this scene's composition" and "Generate a video using a shot of this scene as reference"; and almost all of them "Let AI write from this". Pick one and an empty, connected item appears to the right, waiting for you to generate.

Prompt, model, duration, source references and run status live inside the item's own data. Closing a panel, selecting another item or reopening the board loses nothing; failed jobs settle into a visible failed state with the reason, and failure or cancellation keeps the inputs intact for retry.

Every item shows its run state: queued uses a light primary outline, generating shimmers across the whole item, failed turns red, and cancelled uses a subdued dashed outline. Completion flashes once and settles. Every running item has Stop (generation, voice-over, writing, trimming, abilities); it stops that one job, though a provider may still bill for a remote generation it already accepted. A failed item says "Generation failed" or "Run failed" plus the reason. An item whose job is still alive can't simply be deleted — it asks to stop and delete.

Naming items#

The line above each item shows its name, or its type ("Image", "Video", "Group"…) while it has none. Double-click that line to rename in place, or select the item and use Rename in its action bar: Enter or clicking elsewhere confirms, Esc cancels, and the Enter that picks a candidate in an input method never confirms early. Clear the name to go back to the type. Names are up to 120 characters, and a rename is a single undo step. Board search, list thumbnails, duplicated copies and the AI assistant all use this name.

Referencing media and assets with @#

Press @ in a prompt and a menu follows the caret, listing linked items, then the asset library, then the media library, with filters for All / Linked / Assets / Image / Video / Audio. Only the media kinds the selected model takes are listed; assets (characters, locations, props) always are. What you pick becomes an inline chip in the sentence — "put the person from @creator.png on the street" reads as one sentence. On submit it splits in two: the name stays in the prompt and the media itself is sent as input; how an asset's description and references attach is in Asset library.

Trimming#

Video gets a filmstrip, audio gets its waveform; drag either end to set the range and press Cut it; video can also Grab frame and choose Keep the sound / Drop the sound. For frame accuracy, the number fields on the right are faster. The original is untouched: a trim produces a new asset in a new item.

What an item can do with its content#

There are no separate tool items on a board. Turning an item into something new is something the item itself does: the first four abilities sit directly on the action bar, the rest under ⋯ More abilities:

Item	Abilities
Audio, video	Transcribe asset, Separate voice and background, Reduce noise (strength Light / Medium / Strong)
Video	Video to GIF, Lip sync (audio from a connected audio item, or voiced from a script on the spot)
Image	Split grid, Make it talk
Note, document	Translate (English, Simplified / Traditional Chinese, Japanese, Korean, French, German, Spanish, Russian)
Asset	Fill in angles (all kinds), Generate expressions (characters only), Make it talk (characters only)

Content tools from plugins work the same way, attached to the items whose content they take. Lip sync, Make it talk, Fill in angles and Generate expressions cost money; the two digital-human abilities also need a consent check — see Digital humans.

Split grid cuts a 3×3, 2×2 or storyboard-strip collage into equal tiles: 2 × 2, 3 × 3 (default), 1 × 2 side by side, 2 × 1 stacked, 1 × 3 in a row, 3 × 1 in a column, 2 rows × 3 columns, 3 rows × 2 columns, 3 rows × 4 columns, 4 rows × 3 columns; Gutters can be kept (exact equal split) or removed (up to 6% per side). It runs locally at no cost, and the tiles are laid out to the right in their original grid.

Click an icon and its panel opens under the item; click another and the panel below switches to it; click the same one again to close it (Esc or selecting another item closes it too). The top of the panel says what happens to which item, and this item's content is the input: nothing to pick, nothing to connect.

• The bottom bar holds the pick-one settings ("Target language: English", "Engine"); a chip shows the chosen value, or the field name while nothing is chosen; the rest are under Settings, whose button shows a dot while a required one there is still empty. While something required is missing the send button is grey, and hovering it says what is missing.
• When a tool needs another input, connect that item into this one: the row of upstream chips lets you pick which one feeds which field — notes give their text, documents the pinned revision's text, images / video / audio the asset, 3D scenes the scene. Values are read when you run.
• Each ability's settings are stored on the item, so reopening it shows what you used last time; the item's own content is untouched.

Outputs land to the right: every run adds a new column of items to the right of this one, connected to it — images, video and audio in items of their kind, text as a note, structured data as a code-formatted note. Only the tool's results land, not outputs meant for wiring workflows. Earlier outputs stay. At most 12 items per run; the rest merge into the last note. While it runs, a strip along the bottom shows progress and Stop; after a failure, select the item and the strip says why.

Plugin generators that take no board content (a ComfyUI workflow that turns a prompt into an image, Manim / Remotion animations) are another way to fill an empty item: add an empty image (video, audio) item, switch it to that generator in the action bar, and the panel is the generator's form.

A board holds only content-to-content tools. Flow control and data steps — calling a workflow, HTTP requests, text templates — plugin tools that only list, report or upload, and tools that fetch things by an id from another system (importing ComfyUI outputs by task id, the drive by fs_id) belong to workflows and chat.

One thing, one entry. When a plugin tool is the same thing as a generation model — a ComfyUI workflow that delivers images, video or audio is already a model in image and video items — the board keeps only the Generate entry, and "⋯" no longer lists it again: add an image (video, audio) item and pick that model; the result lands in place. When the workflow has several save nodes, placeholders for runs × images per run are laid out before it runs (a ComfyUI workflow's count is how many times it runs, each run as the workflow is saved; Results per run shows how many runs), and Settings gains Results from: all by default (Final result when the workflow only has preview nodes and some are an intermediate step or a control image, so only the last image comes back), or pick one node to get only its output while the other save nodes don't run. A ComfyUI workflow that only returns text (tagging, prompt extraction) isn't a model and stays an item ability. In a ComfyUI workflow's Settings, size takes any width and height you type (such as 768x1024), the checkpoint and LoRA fields show thumbnails, base model and trigger words, and every workflow with a seed, canvas or not (inpainting and the like), can run several times. See Generating with ComfyUI.

Whose connection: a plugin tool runs with the connection of whoever presses run. On a shared board, settings a teammate saved still run with your own connection — if you have none, the panel names the plugin and links to the Plugins page to create one.

Timeline items#

Add → New → Timeline puts a small editing desk on the canvas. Behind it is an ordinary timeline, kept in a project the board creates under its own name, so you can refine it in the editor.

• The top half is a rough-cut preview: the main video track plays in order with the first audio track alongside, with play / pause, "current / total" and mute underneath. Transforms and captions aren't previewed here — they live in the editor.
• The bottom half is a thumbnail strip (width follows duration) with a playhead: click a clip to select it, drag it sideways to reorder (what follows moves along), Delete removes the selected clip.
• Adding clips: connect video / image / audio items and each is appended to the end (audio goes to the audio track; the aspect follows the first clip; disconnecting doesn't remove the clip), or use the + at the end of the strip, which lists this board's media first and can browse the whole library. Other kinds can't connect — you're told a timeline item only takes video, images and audio.
• Action bar: Split at the playhead, Delete the selected clip, Open in the editor, Export. Splitting, deleting, reordering and adding clips are all in the board's undo history, interleaved with your canvas edits.
• Export: the panel shows clip count, length and size; the bottom bar picks resolution (as is / 1080p / 720p / 480p) and quality (High / Standard / Compact), and Settings holds the "AI generated" label (off by default). The finished video lands as a video item to the right; while it exports a strip above the item says so and can stop it, and the preview keeps working.

Document, asset and 3D scene items#

Document items come in two kinds:

• Referencing a note: Add → New → Document places an empty one. Select it and press Choose a note in the action bar to pin a note at one revision (afterwards Choose another note); the footer shows the revision, with Update to latest revision and Open note. Or press Ask AI to write — empty, it writes a new note and references it; referencing a note, it writes that note's next revision (the old one stays). Ask AI to write sees connected images, assets and notes.
• A document asset: drag in a PDF / Word / slide deck / spreadsheet (or pick one from the library) and the item holds the parsed text, with "Original document · read-only" and Open in the library at the bottom. Save as note on the action bar saves the text as a note and switches the item to reference it, after which Ask AI to write is available. See Documents.

Asset items (Add → From your libraries → Asset) show a cover, the name and a kind badge. Connected into a generation item they act like an @; connected into a note or document their description is material. See Asset library.

3D scene items show a whole-scene blockout rendered by the backend (change the scene and the board shows the new one), and the action bar has Edit scene. The panel's How to produce switches between two modes:

• Render references: pick a Shot in the bottom bar (a single shot is used automatically; with none, open the scene and add one) and Render (First and last frame / Camera-move video / Stills and video), then press the round button; the first frame, last frame and camera-move video land as new items to the right, connected to it, and rendering again adds a new column.
• Build a 3D scene from text: connect a script or storyboard (document or note) into the scene item, pick a model, Max shots (6 by default) and the aspect ratio, and it builds one blockout set — figures move as the text says, several cameras cover the same action from different positions, with camera moves written. The built scene replaces the one in this item (the old one stays under 3D Scenes). Add → New → 3D scene places exactly such an empty scene waiting to be built.

Connect a 3D scene item into an image or video item and the generation panel gets a 3D reference chip: pick the shot and Use as — Composition, First & last frames (video) or Camera move (video), limited to what the selected model takes. At generation time the shot is rendered from the scene as it is then, and the prompt notes that it's a blockout to take composition and camera position from; figures in the shot that Play a character from the asset library bring that character's looks and description along. More in 3D scenes & animation.

The AI assistant#

The robot icon in the top-right opens it. It is not a second AI: messages and confirmation cards are the same ones the global assistant uses; each message just carries one extra line saying which board you are looking at. Each board has its own conversations (you start on a new one that is created with your first message). The top-left title is the current conversation; click it to search, switch, or pick one from Elsewhere to continue here. See AI Studio · Every place has its own conversations.

It can read the canvas (get_board) and change it (edit_board): add notes, rewrite text, name items, connect, delete, move things, place a new timeline item and wire items into it, write export settings. Every change goes through a confirmation card — nothing lands until you approve it, and an open board follows along. It submits granular operations rather than a whole canvas, because a model handed the whole thing gets it wrong on any non-trivial board.

Letting the agent use an item's abilities#

The assistant can use an item's abilities for you. It first looks at what is available (list_board_producers — the same list the action bar reads; plugin tools only from your own connections), then stores the ability's settings on that item and says which other inputs come from which connected items, connecting them in the same batch (edit_board's set_form). A broken edit is refused when the card is opened.

Running (run_board_item) takes the same path as pressing the send button yourself, with the approver's own connection: tools that neither cost money nor reach outside run directly; tools that cost money or act outside the app wait for your approval, and the card names the tool, the board and which of your connections it uses. Auto-approval still applies.

Undo#

⌘Z undoes, ⌘⇧Z (or ⌘Y) redoes, and the toolbar has buttons too. A continuous drag collapses into one step, and splitting, deleting and reordering inside a timeline item share the same history. Undo never removes an item that is still running, and after the board catches up with an agent's edit, undo won't paint that edit back over.

Forms and media selection#

Selecting an item opens a translucent, blurred panel above the canvas. Long model names ellipsize; parameters such as aspect ratio and duration open under Settings. The @ menu follows the input and adjusts to available viewport space.

The recording does not invoke a generation model. Eligible @ media inputs depend on the selected model.

Comments and markers#

The comment and flag buttons enter separate, mutually exclusive modes. Escape exits. Their eye buttons independently show or hide all comments or markers without deleting them. Markers are not listed in Add.

In marker mode, click empty space to place a position bookmark, then set its name and jump shortcut. Only one marker editor opens per canvas. Clear a binding beside the shortcut field or remove the marker using the footer action. In comment mode, click empty space to write a comment and mention workspace members with @. Submitted mentions keep their interactive labels. Use the pencil to edit your own comment. Adjacent comment pins remain directly clickable.

Switching modes or undoing preserves the connections between items. Comments and markers are collaboration overlays and do not take part in generation or workflow runs.