Image Transform

Image Transform is an image-in / image-out filter on the generator contract: Image in, transformed Image out, computed by a single inverse-mapping compute kernel — for every destination pixel the kernel asks "which source position lands here?", walks the transform chain backwards, and samples the source with bilinear filtering. One thread per pixel, no geometry, no intermediate passes.

Work happens in centered, aspect-corrected content coordinates (the Grafit convention: origin at the image center, Y up, shorter edge spanning [-1, 1]). The inverse chain runs: translate → anchor → rotation (positive degrees = clockwise) → perspective → skew → scale → un-anchor, then the nonlinear distortions (twirl, bulge, lens distort, wave), then flips, and finally the resulting source position is converted to texture coordinates and passed through the Edge Mode:

The distortion layer, applied after the affine chain:

• Distort — lens distortion: positive bulges outward (barrel), negative pinches inward (pincushion), strength growing with distance² from center.
• Twirl — rotation angle (in degrees, up to ±1080° = three turns) that falls off smoothly from the center to Twirl Radius; outside the radius the image is untouched.
• Bulge — spherize: positive magnifies the center (bulge), negative shrinks it (pinch), within Bulge Radius.
• Wave — two crossed sine displacements (Wave Frequency controls ripple density); animate Wave Phase with an LFO for flag/water motion.
• Chromatic Aberration — samples red, green and blue at slightly different radial scales (up to ±3% at full strength); red spreads outward, blue inward, alpha taken from the center (green) tap — the classic cheap-lens fringe.

The output texture fills the viewport (the transform can move content anywhere), so Width/Height report the viewport-fit size in Grafit units, and the output is premultiplied alpha like every generator. Enable = false or a disconnected source clears the texture, so Canvas never shows a stale frame.

Affine chain

Distortions

Interactions: Scale multiplies with Scale X/Scale Y, so use Scale for animation and the per-axis pair for fixed aspect correction. The anchor is shared by rotation, perspective, skew and scale — moving it to a corner makes rotation swing the image around that corner. Distortions act on the already transformed image, so a rotated image twirls around the (transformed) view center.

• Infinite scrolling background: Image node → Image Transform (Edge Mode = Repeat, Time → Position X) → Canvas — any texture becomes an endless conveyor; Mirror mode makes non-tileable images seamless.
• VJ kaleidoscope warp: Camera Input → Image Transform (Audio FFT → Twirl, Edge Mode = Mirror, some Bulge) → Canvas.
• Water / flag ripple: any generator → Image Transform (Wave Amount 0.3, LFO → Wave Phase) → Canvas; lower Wave Frequency for slow swells, raise for shimmering.
• Fake 3D card tilt: Image node → Image Transform (Mouse Position → Perspective X/Perspective Y, slight Rotation) → Canvas — a poster that leans toward the cursor.
• Retro lens look: Video Player → Image Transform (Distort 0.2, Chromatic Aberration 0.4) → Image Adjustment (vignette, grain) → Canvas.
• Ken Burns moves: slow LFOs on Scale and Position X/Y over a photo, Edge Mode = Clamp to avoid revealing borders.

• The result is an offscreen texture — route it through a Canvas node to appear on screen; chain it with Image Adjustment or Composite freely, each stays one GPU pass.
• Because sampling is inverse-mapped and bilinear, extreme minification (Scale far below 1, or heavy perspective) can shimmer — there is no mipmapping; keep motion smooth or add softness downstream.
• Edge Mode = Transparent plus a downscaled image gives you a floating picture with real transparency around it — mask- and composite-friendly since everything stays premultiplied.
• For pure placement on the viewport (position/rotation/size with 3D depth), the Canvas node's own transform is cheaper; reach for Image Transform when you need skew, perspective, tiling edges, or any of the distortions baked into the texture itself.
• Twirl accepts up to ±1080° — three full revolutions — which at high values with Edge Mode = Mirror produces dense spiral marbling.
• All parameters are plain float/int/bool ports: everything here is animatable per frame from LFO, Audio, Time or Mouse nodes.