CV Temporal Reduce (Background Plate)

Collapses a whole batch of frames to ONE image by reducing each pixel OVER TIME (across the batch) - the standard way to recover the empty 'background plate' of a fixed-camera video. With the MEDIAN, anything that moves occupies each pixel only a minority of the time and averages out, leaving the static scene (pedestrians, cars vanish); mean is smoother but ghosts moving objects, min/max pick the darkest/brightest value seen. Feed a video IMAGE batch (LoadVideo -> Get Video Components); the plate comes out as a BGR uint8 NPARRAY. A LATENT batch ({'samples': [B,C,H,W]}) also works - the same per-pixel reduction runs over the batch axis in latent space and the plate comes out as a FLOAT32 (H,W,C) NPARRAY (latent values are never uint8-quantized), any channel count. Pair it with 'CV Background Model (Update)' (wire the plate into 'mean', set learning_rate 0) to flag whatever differs from the background in each frame. Also a generic temporal denoiser / long-exposure. Best with frames spread across the clip so moving objects sit in different places.