Every pass may bind up to four inputs: iChannel0 through iChannel3.
An input selects:
channel = 0..3
source = another pass, an asset, "keyboard", "music", or "webcam"
output = pass render-target index, default 0 (pass inputs only)
frame = "current" or "previous" (pass inputs only; output 0 for previous)
filter = "nearest", "linear", or "mipmap"
wrap = "clamp" or "repeat"
Use:
shadertoy inspect channels image --json
to see exactly what a pass consumes.
Sampler settings are per input, including when the same source is connected to
multiple channels with different sampling behavior:
[[pass.input]]
channel = 0
source = "spectrum"
output = 0
filter = "nearest"
wrap = "repeat"
nearest is the right choice for exact numerical grids such as FFT stages.
linear enables interpolated sampling, while mipmap uses trilinear
minification and linear magnification. clamp maps out-of-range coordinates to
the edge texel; repeat wraps them. Omitting these fields keeps the existing
linear/repeat defaults.
The same controls are available when editing a project from the CLI:
shadertoy channel set image 0 spectrum --output 0 --filter nearest --wrap repeat
For a multi-render-target pass, use --output 1 (or another valid attachment)
to bind that target without adding another producer pass.
Media channels
--------------
`kind = "video"` references a local `[[asset]] kind = "video"`. Headless
rendering decodes the frame for deterministic shader time through ffmpeg/ffprobe,
updates the channel texture before each rendered frame, and exposes that time in
`iChannelTime[channel]`.
`kind = "webcam"` uses reserved source `"webcam"`. It is available only in live
preview: the browser asks for camera permission, downsamples frames to 320x240,
and sends RGBA frames to the native renderer over the existing WebSocket.
Headless render/test/replay and preview recording reject webcam input because a
live camera cannot be deterministic.