Rasters (images)
A raster is an image: colour stored on a fixed grid of pixels rather than computed by a formula.
What a raster is
A field is a formula that can be asked for a colour at any point in space, sharp however far you zoom. A raster samples that idea onto a grid of pixels, so many wide by so many tall, each cell holding one colour.
The grid has a fixed resolution, which is all the detail it holds at any one moment. Sampled onto a grid does not mean frozen. The grid can be refreshed every frame: a photo or logo you drag in is a still raster, while a webcam, a video, a rendered frame and a feedback loop are rasters that change continuously.
Finite detail
A raster stores a fixed number of pixels, which gives it limits a formula-based field does not have.
Scale it up and it softens. Stretch a 512-pixel image to fill a 4K canvas and the in-between pixels are invented, so edges blur. There is no extra detail inside it.
Scale it down and it is fine. Shrinking throws detail away, which always looks clean.
Keep something procedural for as long as you can, whether it is a shape, a noise pattern, or a field. Procedural things stay sharp at any size. Turn one into a raster when you actually need pixels: for export, for a paint effect, or to hand to something that expects an image.
Baking and sampling
When a procedural field has to become an image, to display it, paint into it, or feed it somewhere that expects pixels, it is baked: the formula is evaluated across a grid and the result saved as a raster. From there it behaves like any other image, with the finite-detail trade-offs above.
It works the other way too. Drop an image onto an input that expects a colour field and it is wrapped as a field that samples the picture, so a photo plugs into anything that fills with colour, a shape fill or a scatter, with no manual conversion step.
A raster carries its position and size in the scene, so sampling it as a field knows which part of space each pixel covers. Move or scale the raster and its sampled colours follow.
A volume or a field can be sliced on the way into pixels. Wire a Transform into Rasterize's Plane input and drag it to move the cut. Projection decides what the slab hands back: First is a flat slice, while Max, Min, Average and Sum accumulate through the depth, with Samples setting how many steps deep it looks.
Every raster also states how its colour and alpha are stored rather than assuming. A frame captured on the GPU is premultiplied, an image decoded from a file on disk is straight, and anything derived from an existing raster inherits the answer instead of guessing. That is what keeps a soft-edged PNG from fringing white when it composites.
Impostors
A raster does not have to lie flat on the canvas. Content can be captured into a raster and kept showing in the right place in a 3D scene, re-baked as things change. That is an impostor. It stays a raster, cheap flat pixels, and reads as 3D because the image is continuously refreshed from the live view.
An impostor stays a raster. A sprite is a point carrying a raster as its texture, so the value on the wire is points rather than an image. Both are made with the Rasterize node. See Sprites, impostors & billboards.
Sampling between pixels
Two settings decide what a read between pixels returns.
Interpolation. Bilinear blends neighbouring pixels for a smooth read, and is the default. Nearest snaps to the closest pixel for a crisp, blocky look, which is what pixel art needs.
Wrap Mode. What happens past the image border: Clamp, the default, holds the edge pixel; Repeat tiles the image; Mirror tiles it flipped so the seams line up.
Set them on the node doing the reading, not on the node holding the picture. Displace carries both and applies both; Fill carries Wrap Mode and hands it to the material that samples the texture. The matching dropdowns on Image and Video are the ones to ignore: both nodes pass only the file path down to the loader, so nothing downstream ever asks what they were set to, and an image sampled straight off one of them reads bilinear and clamped whatever the panel says.
Frame sequences
A single raster is one still image. A Frame Sequence is a stack of them: frames of a video, a flipbook animation, or captured output over time. A sequence can show the newest frame live, scrub to a specific frame, or hand each copy in an iterator its own frame for per-instance variety.