Node types

The Slicy canvas has a small set of node types. Every workflow is a wiring of these.

Prompt node

The text prompt that gets fed into a model. Plain text, no special syntax. Multi-line is fine.

• Connects to: any model node's prompt input.
• Tips: specific is better than vague. "leather notebook on sunlit desk, soft window light, 3/4 angle" beats "a notebook."

Image input node

A reference image to feed into a model that supports it (style transfer, image-to-image, background removal, upscale, etc.).

• Connects to: model nodes that accept image inputs.
• Input formats: PNG, JPG, WEBP, HEIC. Animated formats (GIF) lose animation.
• Size limits: 16 MB per image, max 5800 × 5800 pixels (matches the tightest model limit).

Model nodes

The actual AI generation step. Different models handle different jobs — see /docs/models for the full menu.

• Cost band above the Run button shows what the run will cost (in credits) BEFORE you click. Final cost is computed from the provider's live per-call price + Slicy's 20 % markup, rounded to 2 decimals.
• Run triggers the generation. Output appears as an Image node (or Video / Text node for those model types) downstream.
• Cancel stops a long-running generation before the provider bills you (works on most providers).

Refine

Post-generation tools that don't run a fresh AI call. Free or cheap.

• Cutout — background removal via BiRefNet (local, free).
• Alpha refine — edge cleanup on cutout output.
• Design text overlay — places copy on the image using golden-ratio + saliency-aware placement.

Save

Sends a finished asset to your asset library so it survives across canvases.

• Drag any image / video / text node into the Save tray at the bottom of the canvas.
• Saved assets appear under your project's asset list.
• Free-tier downloads carry a small "made with Slicy" watermark; paid tiers download without watermark.

How nodes connect

• Drag from a node's right edge (output handle) to another node's left edge (input handle).
• A cable appears. Cables are typed — you can only connect compatible outputs to inputs.
• Multiple cables to the same input are NOT supported — break the old cable first.

Run-flow semantics

• Slicy uses dirty-subgraph re-execution (the same pattern as ComfyUI): when you change a value on a node, only the downstream parts of the graph that DEPEND on that value get re-run on the next Run. Unchanged nodes reuse their cached outputs — no double-billing for the same prompt+settings.
• The cache key is (node config + upstream cache keys), hashed. Identical inputs → cache hit → no provider call.