CV Background Model (Update)

One step of temporal background subtraction with a single-Gaussian per-pixel model (MOG2 without the mixture). Given the running per-pixel 'mean' and 'variance' of the scene so far and the current 'frame', it flags pixels that deviate too far from the model as foreground, then updates the model toward the frame. Because cv2's BackgroundSubtractorMOG2/KNN are stateful objects that cannot travel through a graph (and would double-update under ComfyUI's output caching), the model IS the 'mean'/'variance' arrays - thread them node-to-node like 'CV Compose Pose' threads a trajectory, one node per frame. On the FIRST frame leave 'mean' and 'variance' unwired: the model is seeded from the frame and the foreground comes out empty (no history yet), so no success gate is needed. A pixel is foreground when its distance from the mean, in standard deviations averaged over the colour channels, exceeds 'threshold'. For a FIXED-camera clip you can skip the state chain entirely: precompute a background plate once with 'CV Temporal Reduce', wire it into 'mean', set learning_rate 0, and each frame is scored independently against that plate (which is what lets the detector run inside an Inspire foreach - the loop then carries only ONE accumulator, the output frames). Outputs data only - view the foreground with 'Preview CV Array' (image) or turn it into a MASK for 'CV Overlay Masks' / 'CV Array -> Mask'. Note: the model learns everywhere at 'learning_rate', so an object that stops moving is slowly absorbed into the background (ghosting) - lower the rate to remember longer. Works in LATENT space too: 'frame'/'mean'/'variance' accept a LATENT and echo that format, so the whole detector can run on VAE-encoded frames (the foreground then comes out at latent resolution, H/8 x W/8). BUT the variance/threshold defaults are uint8-intensity^2 scaled; latents are ~unit scale, so for a LATENT frame shrink 'init_variance' to ~1 and 'min_variance' to ~0.01 or the model is so tolerant that NOTHING ever flags.