VAE Latent Inspector

The single most annoying failure mode in ComfyUI is the silent one: you queue a prompt, wait, and get a decode that's black, or static, or a fractal of garbage - with no error, no warning, nothing. Half the time the cause is a latent tensor full of NaN or Inf, and the other half it's a latent with a contrast range so crushed or so blown that the decode is meaningless. VAELatentInspectorMEC is the node that catches both, in one pass, and labels the result with a single word: healthy, low_contrast, saturated, or corrupt.

It's from the ComfyUI-CustomNodePacks pack's diagnostics corner, and it's the kind of node that sounds unnecessary until the first time it saves you twenty minutes of staring at a black image.

How it works

It takes a LATENT tensor (the dict with samples that every sampler outputs), computes per-channel min/max/mean/std, counts NaN and Inf elements, and derives a verdict from the numbers:

• healthy - sane range, no NaNs/Infs.
• low_contrast - values are all squished together; the decode will look flat.
• saturated - values pinned at the extremes.
• corrupt - NaN or Inf present. This is the bad one.

The key detail for workflow use: the node passes the latent through unchanged on latent_passthrough, so you can drop it inline into an existing graph - between sampler and decode - and it inspects without disturbing the data flow. No model, no VRAM cost worth mentioning.

Inputs and outputs that matter

• latent (LATENT) - the tensor to inspect.
• fail_on_corrupt (BOOLEAN, default false) - set true and the node raises a ValueError on NaN/Inf instead of just reporting. Use it as a deliberate tripwire in batch runs.

Outputs:

• latent_passthrough (LATENT) - the unchanged input, so nothing downstream breaks.
• info_json (STRING) - shape, dtype, device, per-channel stats, range, verdict.
• verdict (STRING) - the one-word health label.
• nan_count, inf_count (INT) - exact counts, so you know how corrupt.

Where it fits

The classic deployment is a batch/pipeline guard: wire it between the sampler and VAE Decode, turn on fail_on_corrupt, and a flaky run fails loudly at the exact node instead of silently producing garbage frames. For debugging a specific broken workflow, it's faster than guessing - the verdict plus the stats tell you whether to blame the sampler (corrupt), the prompt/CFG (saturated), or the denoise settings (low contrast).

Installing it

Standard pack install:

cd ComfyUI/custom_nodes
git clone https://github.com/Code2Collapse/ComfyUI-CustomNodePacks.git

restart ComfyUI or use Manager → "CustomNodePacks". Pure torch tensor inspection - no extra dependencies.

The one thing to remember

NaN/Inf in a latent is almost never the VAE's fault - it's upstream (a sampler step, a broken custom sampler, an out-of-range operation). When the verdict says corrupt, this node tells you that it happened; fixing it means looking at whatever produced the latent. That's a feature: it converts "black image, no idea why" into "the latent is corrupt, and it was corrupt before decode," which is where the real debugging starts.