Canvas
The Canvas is where a film is planned before anything renders. It is a node graph, not a folder: characters, locations and shots are cards, and the connections between them decide what each generation sees.
Node types
- Image reference — a character sheet, a location plate, or a prop. Generated or uploaded.
- Audio reference — a voice or sound clip used as a reference.
- Video generation — an actual shot. This is the node that produces footage.
- Note — annotations for the crew. Never sent to a model. A note has a type,
and the same text reads differently for each: Beat (plain story moment),
Dialogue (an ALL-CAPS line becomes a character cue,
( … )a parenthetical), Direction (CAM:,LENS —,LIGHT -and friends become tags), Continuity (key: valuerows) and To-do (one task per line, with checkboxes and a done counter). Select a note to change its type or colour from the chip in its header, resize the text with A− / A+, or drag the bottom-right corner. Double-click the body to edit. - Section — a labelled frame for grouping. Purely organisational. Drag any of its four sides to stretch it — no need to select the section first — but it never shrinks past the nodes inside. Select it to rename, recolour, collapse it into a single pill (its nodes and their connections hide until you expand it) or delete the frame; deleting keeps the nodes on the canvas, and Undo brings the frame back.
Presets — a model and its settings, saved together
Every generation node carries a model plus its parameters (resolution, ratio, duration, audio for video; ratio, resolution, quality for images; kind, model and length for audio). Picking that combination again for every node is the tax this removes.
The preset button sits on the model row, next to compare. Save what you are using now under a name, and one click restores the whole combination later. Presets live on your account, not on a project, so they follow you into the next film. Mark one default and every new node of that type is born with those settings already applied; the default always sits at the top of the list.
The button is highlighted only while the current settings still match a saved preset — change one value and the highlight drops, so it never claims something the card is not doing. Hover it to see which preset is in play.
Looking at a shot up close
The ⛶ button on any thumbnail opens the full-size viewer. Pinch on a trackpad
(or hold ⌘/Ctrl and scroll) to zoom into the image itself rather than the app,
scroll with two fingers to move around, and drag to pan. Double-click toggles
between fit-to-screen and actual pixels; + / - / 0 / 1 do the same from
the keyboard, and Esc closes.
Connections are what make consistency work
Connecting a character sheet to a shot means that character's appearance travels into the generation. That is the whole mechanism behind "the same person across every shot" — there is no hidden memory, just references you can see.
Connect a location plate the same way to keep a set consistent. A film's look is not a node — each prompt carries its own style wording, and the agent keeps that wording identical across shots.
An edge to a node that already generated is locked: changing what a finished take was made from would make its record a lie.
Video into video — Reference or Last frame
Every node has one input on its left edge and one output on its right. Drag from an output and drop anywhere on the target card. When the source is a video shot, the card splits into two drop zones as you hover — release on the one you want:
- Use as reference (bottom half, the default) — the whole clip goes in as a video reference for motion, look and character. Several clips can feed one shot. Click a video thumbnail in the inspector (or Trim clip… in the badge menu) to use only part of the clip; the source stays untouched and you are charged for the trimmed length.
- Continue from last frame (top half) — the final frame of the source's selected take becomes this shot's start frame, so the cut continues exactly where the last one ended. A shot can have only one of these; picking another replaces it. While the link is in place the shot is in first / last frame mode and nothing else is sent. If the source has no take yet, or its last frame is still being extracted, Generate waits rather than spending credits.
Video-to-video lines carry a small badge at their midpoint (↪ last frame, ◎ reference). Click the badge to switch modes, trim, or disconnect. Clicking the line itself disconnects it — a toast with Undo puts it back.
A take that already generated shows the mode it was made with, so the first / last frame pill reflects that record rather than your live settings. Pressing it starts a new take from that record and flips the mode there.
Deleting a node that other clips reference asks first, and says which clips lose that reference — a shot regenerated without its character sheet comes back looking different, and that is hard to notice after the fact. The node's media moves to the media Trash, where it can be restored for 30 days.
References inside the prompt
References attached to a shot can be mentioned in the prompt by clicking their thumbnail, which inserts a mention at the cursor. That is how you say this character rather than describing them again.
Takes and the settings card
Every generation lands as a new take on the same node, and nothing is ever overwritten. The takes row shows them as tiles; click one to look at it, double-click (or Set as best) to make it the take other nodes reference.
A take you are looking at is a record: the card header reads settings · T2 and
everything on screen is what that take was made with. Records are never
overwritten — but you can still start typing. The moment you change anything
(the prompt, a setting, a reference, a frame slot), an empty card appears at the
end of the row with that take's settings copied into it, and your change lands
there. The header switches to new take · copied from T2 and the button becomes
Generate.
New take is still in the footer for when you want to run the same settings again without changing anything. It is free and does not generate on its own. The empty card is saved with the node, so it is still there after a reload or on another device. You can dismiss it with the ✕, or click back to a finished take to compare and then return to it without losing what you typed. If you change something while looking at a different take, the empty card restarts from that take instead — a message offers Undo if that was not what you meant.
Removing a reference with its ✕ only leaves it out of the next take — the connection on the canvas stays, and a small pill under the strip lets you use it again. To actually disconnect two nodes, remove the line on the canvas.
Settings per shot
Each generation node carries its own model, resolution, ratio, length and seed. They live in the settings card above the references, and the credit estimate on the Generate button follows them immediately.
Storyboard
The dock at the bottom groups shots into scenes so a long canvas stays readable. Scenes are labels over the same nodes — moving a shot between scenes does not change what it references.
Working on mobile
On narrow screens the canvas plane is replaced by a list of nodes with an inspector. Reading, checking takes and starting generations all work; laying out the graph is a desktop job.
The list loads the first 12 nodes and adds more when you tap Show more. This keeps a large project from pulling its whole media library into a phone browser at once, which is what made big canvases reload themselves on mobile Safari.
Last updated 2026-09-15. Prices, model specs and credit costs on this page are read from the live catalogue at build time.
