Mask & Latent

The Mask & Latent node packs fourteen operations onto one dropdown, split between the two kinds of data that live between "image in" and "image out" in a ComfyUI graph: masks (the grayscale selection maps that drive inpainting and compositing) and latents (the compressed tensors your sampler actually denoises). The mode dropdown covers mask-from-color, mask-from-depth/luma, invert, erode, dilate, blur, combine (union/intersect/diff/xor), and inspect on the mask side; latent inspect, math, noise inject, smart upscale, a NaN guard, and pad/crop on the latent side. Outputs are mask, latent, and info (a DICT).

The two modes I'd actually install this pack for: mask_from_depth and latent_nan_guard. Depth maps (from MiDaS, ZoeDepth, or any of the depth estimators floating around) encode "close is bright, far is dark." mask_from_depth thresholds that map by a min_value/max_value range - default 0 to 0.5 grabs the foreground - and hands you a selection you can feed straight into inpaint or a denoise-masked KSampler. It's the same trick that powers a lot of the background-removal and depth-driven workflows in the ecosystem, done in a single node. And latent_nan_guard is the one to reach for when SDXL FP16 spits out NaNs and your VAE decode renders a pure black image: set on_nan to raise to fail fast with a clear message, or zero_out/clamp to recover and keep going.

How it works

Standard pack architecture: mode picks the operation, irrelevant widgets hide, unused outputs pad. Key inputs:

• mask_from_color: image plus r/g/b and tolerance - mask is 1 where pixels are within tolerance of that RGB.
• mask_from_depth: image, min_value, max_value, invert.
• Erode/dilate use kernel_size; blur uses radius; combine takes two masks a/b and a combine_op.
• Latent ops take a latent: latent_math (add/subtract/blend with weight), latent_noise_inject (strength + seed), latent_upscale_smart (scale - picks bilinear under 2×, bicubic above), latent_pad_crop (target_h/target_w/fill_value).

The latent ops have a detail worth knowing: they preserve the auxiliary keys in your latent dict - things like noise_mask, which inpainting pipelines attach to the latent. Older versions of the pack silently dropped those and broke masked inpainting; current releases shallow-copy the dict and only swap the samples. latent_pad_crop also deserves a shout-out: it pads or center-crops a latent to a target size in latent space, so you can align two latents for compositing or snap to an SDXL bucket without decoding to pixels and back - that saves a whole VAE roundtrip.

Installing it

One node from the pack - install the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/IxMxAMAR/ComfyUI-Utility-MegaPack

Restart ComfyUI or use ComfyUI Manager → "Utility-MegaPack". No model files; dependencies are the usual light set.

Where people get tripped up

Erode and dilate force odd kernel sizes under the hood - an even kernel_size used to produce asymmetric padding and crash downstream nodes with off-by-one shapes, so the pack just bumps even numbers up. Don't fight it. mask_combine requires a and b to have identical shapes or it raises; if they don't match, resize first. And the NaN-guard raise option is deliberately loud: a NaN latent decodes to a black image no matter how much GPU you spend, so stopping is the feature. One aside: the info output is where all the inspect modes land - mask_inspect gives you coverage percentage, bounding box, and centroid, which is handy if you're auto-cropping to a subject instead of eyeballing it.