How does the Martini canvas work?

Every node on the canvas is one model call. An Image node generates an image when you click Run. A Video node generates a video. An Audio node produces sound. A Text node runs a text or script model. Each node has a model picker, a prompt field, parameters specific to that model, an output preview, and a Run button. Nodes are independent — running one does not run the others unless they are connected and you trigger a downstream chain.

The node panel shows the credit cost for the next run before you click. This is by design: you see the cost in advance and can adjust resolution, duration, or batch size to control spend. Failed runs on a node do not break the rest of the canvas; you fix the inputs and re-run that node alone. The canvas is composable, not monolithic.

Each node has input and output ports. Drag a line from one node's output to another node's input to chain them. The connection turns green when valid (the upstream output type matches what the downstream input accepts), red when invalid (a video output cannot drop into a text-only input). This is the main authoring action on the canvas — drawing the workflow.

When you run a downstream node, Martini reads the upstream output as input. Image-to-video reads the image; video-to-lip-sync reads the video and an audio source. If you re-run an upstream node, the output changes, and downstream nodes use the new output the next time you run them. There is no global Run button — each node runs on its own click. This keeps re-runs cheap and intentional.

Click and drag empty canvas space to pan. Scroll the wheel (or pinch on a trackpad) to zoom. Hold shift and click multiple nodes to multi-select; shift-drag to box-select a region. Selected nodes can be moved together, copied, deleted, or grouped. The mini-map in the bottom-right shows the full canvas and lets you jump to distant regions on a large project.

Keyboard shortcuts: Cmd/Ctrl+C to copy selected nodes, Cmd/Ctrl+V to paste, Delete to remove, Cmd/Ctrl+Z to undo, Cmd/Ctrl+A to select all. The undo history covers node creation, deletion, parameter changes, and connections — but not generation outputs (those are immutable once produced). Lost a great output? Use the generation history panel to retrieve it.

A chat-style AI tool gives you one input and one output per turn. To build anything multi-step, you copy outputs forward, paste into the next prompt, and re-run. The canvas removes that loop. You see every step of the pipeline at once, you re-run individual steps without recomputing the rest, and you branch alternatives by duplicating a node and changing one parameter.

For complex briefs — multi-shot videos, voiceover-driven explainers, character consistency across scenes, ad creative variants — the canvas is the difference between thirty browser tabs and one workspace. For a single quick generation, the canvas may feel heavy; for everything beyond that, it is the unlock.