What is LINK?

LINK is a node-graph compositor. Drop nodes on a canvas, connect them by channel type, and run the graph — LINK orchestrates the work across local compositing operations, cloud vendor models, and your own ComfyUI workflows. Same canvas, same wiring, regardless of where each node actually runs.

Open canvas →

Quick start

Five minutes from a clean canvas to a finished render:

Drop a source. Drag a vendor node from the left palette — flux or seedance are good first picks. It lands wherever you release.
Connect. Drag from an output port (right side of a node, coloured by channel type) to an input port on another node. Edges follow the source port's colour so channel flow is visible at a glance.
Adjust. Multi-port nodes like merge show port labels (fg / bg / mask). Hover any port to see what it carries.
Run. Click Run graph in the topbar. Per-node state updates in real time; the log drawer at the bottom shows progress per node.
Preview. Open the Viewer panel from the topbar to see the focused node's output. Switch between channels with the R·G·B·A buttons.

Hotkeys

Tab

Open the quick-create overlay. Type to filter, ↑ ↓ to navigate, ↵ to spawn at the cursor.

Delete / Backspace

Remove every selected node. Edges that touched a removed node go with it.

⌘A / Ctrl A

Select every node on the canvas.

Shift + click node

Toggle the node in the selection — build a multi-selection one node at a time.

Shift + drag empty

Marquee select. Every node whose box intersects the rectangle joins the selection.

Drag empty canvas

Pan the world. Scroll wheel zooms around the cursor.

Drag a selected node

Moves the whole selection. Relative spacing is preserved through snap-to-grid.

Right-click empty canvas

Grid style (dots / lines / none), snap-to-grid toggle, tidy layout, fit view, reset view.

Drag from output port

Pulls an edge towards the cursor. Snaps to nearby input ports of compatible type.

Esc

Close the quick-create overlay or the right-click menu.

Node types

Local comp · `link_comp`

Pixel-math operations that run on Lenscowboy infrastructure: premultiply / unpremultiply, resize, A-over-B merge, format conversion, color-space transforms. Fast, cheap, deterministic. Caches aggressively — rerunning the same graph with the same inputs is instant.

Vendor · `vendor`

Generative AI models from external providers — image, video, and audio. Text-to-video, image-to-video, video restyle, image generation, music, voice. Cost depends on the model and duration; the run dialog shows estimates before you commit.

ComfyUI · `comfyui`

Your own ComfyUI workflows, run on managed GPU hardware. Upload a workflow JSON, declare its inputs and outputs, and LINK treats it as just another node — with caching, retry, and the same canvas wiring as everything else.

PAL · `pal`

Source node carrying a multipass render from a PAL scene — beauty plus auxiliary passes (depth, normals, motion, mask). Use the “Send to LINK” button in PAL's topbar to drop a fully-populated pal_render node on a new canvas.

Local · `local`

Nodes that execute on your machine — never the cloud. Reading EXR files from disk, pulling assets from ShotGrid or ftrack, generating proxies. Marked with a green L badge in the upper-right corner. Requires the LINK Local Worker installed and paired to your account.

The Viewer panel

Toggle the Viewer from the topbar. Side dock parks it on the right (360px); Top dock parks it above the canvas (320px). The viewer shows whichever node is focused — before a run it shows the node's primary input; after a run it shows the executed output.

Channel buttons (RGB / R / G / B / A) isolate single channels for inspection. Useful when a comp's misbehaving and you need to see whether the alpha actually went where you thought it did. Below the preview, an info strip shows the asset's URI, channel layout, color space, and content hash.

Color management

Every output is tagged with its color space; every transform respects it. The defaults follow OCIO 2.4 (ACES 1.3) — comp math in ACEScg linear, output baked through an sRGB view transform. Adjust the universal baseline at Settings → Color; override on a per-graph basis by dropping a color_transform node anywhere in the chain.

Workflows

1. Quick T2V

Drag seedance or any T2V vendor onto the canvas. Set the prompt in the node panel. Run. Preview in the viewer.

2. Source image → video

Drag an image generator (flux, midjourney), drag kling or another I2V vendor, connect the image output to the I2V input. Run.

3. Two-layer composite

Drag two image sources. Add premult nodes after each (alpha needs to be premultiplied before merging). Add merge, wire FG and BG into its respective ports. Optional: add a mask source into merge.mask. Run.

4. Format conversion & color round-trip

Drag format to bake out PNG / JPG / EXR; drag color_transform to change colour space (sRGB ↔ ACEScg, etc.). Both are local-comp and run fast.

5. PAL handoff

In PAL, render a scene with AI Render. Click LINK in the PAL topbar — a new LINK canvas opens with a pal_render node pre-populated. Drag downstream restyle / comp nodes; the multipass auxiliaries are available as separate outputs.

What happens on errors

Cold start — the first run of a graph on a new tenant can take a minute while infrastructure warms up. Subsequent runs are fast.
Vendor errors — surfaced inline on the node + in the log drawer. The retry policy will automatically re-run idempotent steps; non-idempotent ones (most vendors) fail loudly so you don't get double-charged.
Cancelled — clicking Run again while a graph is mid-flight cancels the previous run. The log marks cancelled nodes; cache entries from earlier successful nodes survive.
Local-kind node without local worker installed — a node marked with the green L badge will error with a clear message pointing at the install instructions. The cloud knows nothing about your local files.

Where things go

LINK outputs land in your Drive under the project folder — same place PAL renders and Pipeline outputs go. Each graph run produces a deterministic content hash; identical inputs produce identical outputs (caching is content-addressed, not URI-addressed).