NLMeans

The texture defender. It searches the frame for fragments similar to a given one and averages them — so detail survives while noise goes away.

What it does

It splits the frame into small squares (patches) and, for each one, searches its neighborhood for similar squares. The similar ones found are averaged: noise in them is random and is canceled out by averaging, while the underlying structure matches — so texture stays texture. Unlike the block-based algorithms, it searches across the whole frame rather than within a limited block.

When to pick it

• Textures and fine detail matter: skin, hair, fabric, foliage, brick.
• HQDN3D or FFTdnoiz eat detail, but you can't afford to skip the cleanup.
• The frame has motion — the algorithm is spatial, so motion doesn't bother it.

When not to pick it: you need speed (this is one of the slowest algorithms), or the footage is simple and speed matters more than quality.

Parameters in plain words

• Strength — the cleanup strength. 0 — off; 1 — gentle; 15+ — very strong, with a risk of blurring.
• Patch Size — the size of the square being matched (an odd number, 7 by default). Bigger — more aggressive averaging, better noise handling, worse detail.
• Research Size — the size of the area searched (15 by default). Bigger — better quality, but much slower: the search is quadratic in area.
• Chroma Patch Size / Chroma Research Size — the same for color. 0 by default (off): color cleanup in NLMeans is a second full search, and it is expensive. Turn it on if color noise is strong and other algorithms don't handle it.

In Simple, the Denoise slider moves Strength, Balance moves the Patch and Research sizes (further left is faster, further right is better quality), and Chroma is hidden for NLMeans: in Simple mode the color pass is not enabled.

How it works inside

NLMeans compares patches — squares around pixels — rather than the pixels themselves: two pixels are "similar" if their neighborhoods are similar. This makes it possible to tell noise (random, unlike anything else) from texture (similar to other parts of the same texture). The averaging weight is inversely proportional to the distance between patches: the more similar they are, the stronger the contribution. That is exactly why the algorithm keeps detail that ordinary blurring considers "noise".

If it didn't help

• Very slow → smaller Research Size, Balance to the left, lower Strength; or use FFTdnoiz / BM3D instead.
• Blurred what it shouldn't → smaller Patch Size, lower Strength.
• Color noise remains → turn on Chroma Patch/Research Size, or ATAdenoise in the neighboring slot.
• Luma noise remains → higher Strength or a second slot (BM3D).

Next — FAQ or the next algorithm: FFTdnoiz.