📟 Gimbal Latent Telemetry (LAMNr OOD)

GimbalLatentTelemetry is the most "research-grade" node in the pack: it scores your latent with four statistical metrics and tells you how far out-of-distribution it is. If GimbalDiagnostics is the check-engine light (min, max, mean, std), this is the full diagnostic report from the lab - log-likelihood, Mahalanobis distance, total correlation, and geodesic distance. You wire a latent in, read the numbers, and decide whether what you're about to sample is in the healthy zone or wandering in the weeds.

How it works

Inputs: latent (required) and three optional knobs - comparison_latent (the target for the geodesic comparison; if absent and your batch has more than one sample, it measures from the batch centroid), subspace_rank (the low-rank cap for the Mahalanobis estimate, −1 = all SVD components, 0 = isotropic-only baseline), and bandwidth (the Gaussian bandwidth for the total-correlation estimator, 0 = auto via median pairwise distance). Leave the optional ones alone until you know why you're changing them.

Outputs: latent_passthrough (it passes your latent through untouched - this node never rewrites it), then four FLOATs: mean_log_likelihood, mean_mahalanobis, mean_total_correlation, mean_geodesic, plus advanced_telemetry (the dict with the per-sample breakdown). The metrics are the classic OOD toolkit: Mahalanobis distance measures how many standard deviations you are from the distribution's center along its principal axes, log-likelihood scores how probable the latent is under the fitted Gaussian, total correlation measures how entangled the channels are, and geodesic is the spherical distance between two latents.

When you'd reach for it

Honestly? This is a niche node for a niche audience. If you're systematically testing steering strengths or building a pipeline where you need to quantify "is this blend still on the manifold" instead of eyeballing it, this is the tool - pair it with GimbalLatentMath's metric ops for a fuller picture. For a beginner, the pragmatic use is: run it before and after a big steering move, and if Mahalanobis jumps by an order of magnitude, that's a sign to back off the strength or run a GimbalLatentStabilizer before sampling. The trend matters more than the absolute number - there's no universal "healthy" cutoff, and the pack's docs don't give you one.

Installing and gotchas

Standard pack install: ComfyUI Manager → search "Gimbal-comfy" → install → restart, or git clone https://github.com/FormAndNoise/Gimbal-comfy into ComfyUI/custom_nodes. Dependencies are torch/numpy/pillow, already bundled; nothing downloads. (README's clone URL is stale - FormAndNoise/Gimbal-comfy is live.)

The trap is over-reading the numbers. A single latent has no meaningful mean_geodesic without a comparison target - the metric only starts to mean something with a batch or a comparison_latent. And total correlation with bandwidth = 0 auto-selects via median pairwise distance, which on a tiny batch can be noisy. Treat every number here as a relative signal, not an absolute verdict. It's a genuinely competent implementation of standard OOD math - but it's the kind of tool that rewards users who already know what Mahalanobis distance is, and the pack's near-zero community presence means nobody's writing friendly explainers for you.