Latent Anomaly Mask 🚩

Latents are the compressed representation the diffusion model actually works in - 8x smaller per side than the image, 4 channels for SDXL, 16 for Flux. You can't see them, and that's the problem this node exists for. Latent Anomaly Mask runs a statistical scan over a sampled latent and flags the spatial regions that are statistical outliers - the hot spots, the corrupted blobs, the parts where something went wrong before you've spent a VAE decode finding out. Think of it as an X-ray for your latent, showing you the injuries before you develop the film.

How it works

For every spatial position it computes the local mean and standard deviation in a sliding window of window_size pixels, then compares the local value against that neighborhood. Anything more than std_threshold standard deviations away gets flagged. That's a per-region check, not a global one - a latent can be perfectly healthy overall and still contain a small pathological patch, which is exactly what a global mean/std check would miss.

It returns three things:

• mask - highlights the flagged regions. Wire this into an inpainting workflow after decode, or use it with Image Mask Composite to paint over the bad area.
• corrected_latent - with auto_correct on, the flagged positions get softened toward the local mean by correct_strength. A light touch (0.3–0.5) can smooth a glitchy region without flattening the whole latent.
• report - a text summary of how much of the latent got flagged.

Inputs that matter

Only a few are worth touching:

• window_size (3–31, default 5) - how big a neighborhood defines "local". Bigger windows call more things anomalous because the reference region is broader.
• std_threshold (default 3) - how many standard deviations before it's "anomalous." 3 is a reasonable start; raise it if the mask lights up like a Christmas tree on healthy images.
• auto_correct / correct_strength - whether to actually fix, and how hard.

When you'd actually reach for it

This is a debugging and QC node more than a creative one. The best use is in a batch run: throw it after the KSampler, read report, and let a gate decide whether to decode a latent that came out looking like a glitch. It pairs naturally with the pack's Latent QC Gate (the automated pass/fail version of this idea) and Latent Histogram (which shows you the distribution shape). If you're doing unattended batch generation, a node that catches the corrupted samples before the VAE wastes time on them is worth its weight.

Install

Standard OmniNodes story - no extra dependencies, because latent math here is pure NumPy/PyTorch, both bundled with ComfyUI:

cd ComfyUI/custom_nodes
git clone https://github.com/TensorVizion/OmniNodes

Restart, then look under TensorVizion/Latent in the node search. ComfyUI Manager (search "OmniNodes") works too. After cloning, check the terminal for [OmniNodes] ✅ Loaded lines so you know the file registered.

Troubleshooting

• Everything is flagged. Drop std_threshold is the wrong move - raise it. If 3 flags half the latent, either the latent is genuinely broken or you want a 4–5 threshold.
• Nothing is flagged but the image looks bad. The anomaly is global, not local - a uniform wash won't trip a local outlier check. That's a job for the QC Gate's near-blank check instead.
• The mask and corrected_latent disagree with each other. They're separate outputs doing separate jobs. mask is diagnostic; corrected_latent is what you actually feed downstream. Don't feed the mask into a decode expecting it to be an image - it's a mask.

It's a genuinely useful diagnostic in a pack that's mostly about giving you eyes inside latent space. A beginner's honest take: use it when something's visibly wrong and you want to know where, not as a permanent fixture in every workflow.