The lines on the DOTV canvas are not decoration. A line says one thing: the one on the left is an input to the one on the right. So a glance at a canvas tells you how the AI video was made — which image referenced what, which frame a clip was built from. This section covers what a line actually feeds to the next node, how to wire one up, how to build the next step while you're at it, and how to cut one.
A line only does its work at the moment of generation. If the image on the left is regenerated later, the one on the right does not follow on its own — to bring it up to date, hit generate again.
A line is not a "relation", it is material. When the node on the right starts generating, everything wired into it goes to the model together. Which is why what can feed what is not open-ended — an effect phrase pasted into an image request does nothing at all, and a line that is wired but unused only misleads you, so it is blocked at the moment you draw it:
The line itself runs square and steers around cards: two cards set at an angle never produce a wire that cuts diagonally across some other node. When one side of a node carries several lines, the openings spread a little above and below each other and the middle stretch of each wire takes its own lane, so the lines run parallel the whole way instead of merging into one thick line somewhere.
| This node | Only takes |
|---|---|
| Image | Images (as a reference image), style samples |
| Video | Images (as the first frame / a reference), videos (as a reference video in omni-reference mode), effect samples |
| Text / script | Text and scripts (as the source text, written on top of it), images, video (the model reads the picture or the footage) |
| Audio | Text, scripts |
| Smart edit | Video |
| Frame study | Video |
| 3D previz | Images (as the panoramic backdrop) |
| Style / effect sample | Nothing at all — a sample dropped in from the libraries has no upstream |
The audio row is the exception: a dub reads the text in its own input bar, and a script wired into it never reaches the request — that line only records where the dub came from.
The lines from style and effect samples go further: wiring one up is applying that look — its cover appears on the input bar right away (see 10.4), and on a node that has not generated yet the ratio switches over with it. Only one of each can be attached at a time: apply a new one and the old line disappears by itself.
A node can take several sources, and the order does not matter — in the picture above, the character sheet and the set image were fed to the same shot together.
Groups are not on this table: lines run between nodes, and there is nothing to attach to a group's bed (see 9.2).
A pair that cannot be wired behaves differently on the two routes: dragging a line is blocked, with a message naming what does fit, read straight off the table above (only an image node or a style sample can feed an image node); ⌘L simply does nothing — both nodes are still selected, and nothing happening is the answer: this pair cannot be wired.
Three ways, same result:
The "+" on the left runs the other way — dragging from it adds a reference to this node.
When one end is an empty video that has never produced anything and the other is a video that already has its result, the gesture does not decide the direction: the finished one is always the source, and the empty video takes it as a reference video (the same for dragging and ⌘L).
The two openings follow the table above: a node that can feed nothing has no "+" on the right (audio, smart edit, 3D previz, frame study), and a node that takes nothing has none on the left (style and effect samples, and text nodes you wrote by hand).
Click the "+" instead of dragging, and a menu opens:
Pick a type and the new node is created on the right and wired up — no building a node and then coming back to connect it. Dragging a line and letting go on empty canvas opens the same menu; the only difference is that the new node lands where you let go.
The "+" on the left opens the same menu under a different title, Generate a node to reference: the new node is built on the left and wired into the current one.
The menu always holds the same types, and the ones you cannot reach from here are greyed out (open it on a style sample and only Image is live; on an image node it is image, video and script) — the positions never move, so you remember which one is where, and the greyed rows say plainly that this route does not lead there.
Delete a node and the lines attached to it go with it, without asking. ⌘Z brings back the node and the lines together.
Rest the mouse on a line and it thickens, with a pair of scissors in the middle; click them and it is cut. Both nodes stay where they are, they just stop being inputs to each other.
You can also delete it from the referencing node's composer: each reference thumbnail in that row has an ×, and removing one is the same as cutting that line.
Cutting the line from a style or effect sample is a bigger deal: it un-applies the look, and the style comes off the input bar with it. Deleting the sample node, or clicking the × on the cover in the input bar, does exactly the same thing.