Composite

Composite renders on two GPU compute passes per frame. The first composites every connected layer into an internal texture at the node's own exact pixel size, bottom-to-top in layer order (Image 1 is the bottom layer, Image 8 the top). The second "present" pass letterboxes that finished composite into the viewport-sized output texture, aspect preserved with transparent bars — so a downstream Canvas shows the composite at its true proportions, and a chained Composite can read the content rectangle back through the standard Width/Height contract.

Size Mode decides the internal canvas resolution (clamped to 16–8192 px per side):

Compositing math follows the W3C spec on straight (un-premultiplied) colors: each layer's texture is un-premultiplied, blended against the accumulated backdrop, then combined with source-over alpha. Coverage per pixel is source alpha × Opacity × mask. The final write re-premultiplies, and Background Color seeds the accumulator (default fully transparent).

Each layer is re-fitted onto the composite canvas by inverse-mapping its Fit mode against the composite's aspect — and crucially, the fit works on the layer's content rectangle (its declared Width/Height in Grafit units) rather than on the padded viewport texture, so a small shape doesn't get treated as a full-frame image:

A layer's optional Mask is sampled in composite space and always covers aspect-preserving (fill, never stretched — a silhouette keeps its proportions on any canvas). Mask coverage comes from the mask's premultiplied luminance: white opaque areas pass the layer through, black or empty areas block it.

Blend selects how the layer's straight color s combines with the backdrop b beneath it (the blend result is mixed toward the plain source where the backdrop is transparent, per W3C):

Layer port groups appear progressively: Image 1 is always visible, Image N appears once Image N−1 is connected, and a layer's Mask, Blend, Opacity and Fit ports only appear once its Image is connected. Likewise, Width/Height show only in Custom Pixels mode and Aspect W/Aspect H only in Aspect Ratio mode. With Enable false, no layers are composited — only the transparent/background canvas is produced.

Canvas

Per layer, N = 1…8

Outputs: Image (the composited texture), plus Width/Height reporting the composite's content size in Grafit units — which is exactly what lets a chained Composite or a Canvas place it at its true aspect ratio.

Practical interactions: Opacity N and the mask multiply together, so a 50% layer under a mid-gray mask reads at 25%. Blend modes operate against everything already accumulated below, not against a single layer — reordering connections changes the result. Original Size maps source pixels 1:1 onto the canvas, which means a small source lands small even on an 8K canvas.

• Fixed-resolution delivery frame: several generators → Composite (Size Mode = Custom Pixels, 1920×1080) → Canvas — build the show at broadcast resolution regardless of window size.
• Vertical social cut: the same graph → a second Composite with Size Mode = Aspect Ratio, Aspect W 9 / Aspect H 16, layers set to Fill — an instant portrait crop of the same scene.
• Masked reveal: Video Player → Image 1; Text generator → Image 2 with a Shape generator into Mask 2 — the text appears only inside the moving shape.
• Glow stack: base render → Image 1 (Blend = Normal), the same render blurred → Image 2 (Blend = Add, Opacity 0.6) → Canvas.
• Duotone / grade over footage: Camera Input → Image 1; a flat Color generator → Image 2 with Blend = Overlay or Soft Light at partial opacity.
• Chained composites: Composite (background plate) → Composite (Image 1 = plate, Image 2 = foreground elements, Blend = Screen) → Image Adjustment → Canvas — group layers logically and grade the whole stack at once.

• The output is still an offscreen texture: it needs a Canvas node to reach the screen. Layer ordering happens inside Composite (Image 1 = bottom); the Canvas layer system is separate and only orders Canvas nodes.
• Prefer Composite over many Canvas nodes when you want a fixed pixel canvas, per-layer masks, or blend modes richer than Canvas's four — and when you want to grade or transform the whole stack afterwards as one image.
• Masks read luminance from premultiplied pixels: an empty (transparent) mask region blocks the layer entirely, exactly like black. A generator that outputs a white shape on transparency is a ready-made matte.
• Fit works on the content rectangle each source declares through its Width/Height outputs, so a small generated shape stays small under Fit/Fill instead of being blown up to the full padded texture — use Stretch deliberately when you actually want distortion.
• Very large Custom Pixels canvases cost fill rate on every frame; 8192 is the hard ceiling per side, but a 1080p or 1440p canvas is usually the sweet spot for live work.
• Related nodes: Image Adjustment for grading the composited result, Image Transform for moving/warping a layer before it enters a slot, and Canvas for final placement in the 3D output scene.