BM3D

Block-based quality. It groups similar blocks of the frame and averages them as whole groups — the reference among spatial algorithms.

What it does

For each block it searches the whole frame for similar blocks, stacks the ones it finds into a group and cleans the group jointly: noise in the group averages out, while the common structure of the blocks matches, so that is what gets reinforced. The result is a clean picture with living textures where the other algorithms gave up and painted plastic long ago.

When to pick it

• You need high quality on a single frame and time is not limited.
• Medium and heavy noise.
• As a second slot after a fast pass (HQDN3D → BM3D).
• Footage with textures that have to survive.

When not to pick it: drafts and real time (the slowest of the spatial algorithms), or very simple footage (overkill).

Parameters in plain words

• Sigma — how strong the noise must be to count as noise. It sets how aggressively the algorithm cleans; setting it too high blurs detail.
• Block Size — the block size (4–16, 8 by default).
• Block Step — the step of the block grid (1–64, 4 by default). Smaller — denser coverage, better quality and slower.
• Group Size — how many similar blocks are taken into a group (1–256, 8 by default). More — better noise suppression.
• Search Range — the radius in which similar blocks are searched for (9 by default). More — better quality, exponentially slower.
• Match Step — the step at which positions are traversed during the search.
• Match Threshold — how similar blocks must be to get into a group. Lower — stricter selection (fewer "foreign" blocks in the group).
• Hard Threshold — the hard zeroing threshold in the second stage (2.7 by default).
• Estimation — Basic (one stage) or Final (two stages: an estimate first, then a polish based on it). Final is better quality and twice as expensive.
• Use Ref — use the reference frame (the previous one) as an additional group member. Works together with motion compensation.
• Planes — which planes to process (luma and color).

In Simple: Denoise moves Sigma, Balance moves Block Size, Block Step and Search Range (further left is faster, further right is better quality).

How it works inside

A two-stage scheme: first a group of similar blocks is assembled for each block and a "clean" block is estimated from it, then in the second stage that estimate is used as an expert filter for the final smoothing. Block similarity is the main source of quality: noise in blocks is independent, so averaging over a group cancels it in proportion to the square root of the group size, while matching structure is only reinforced.

If it didn't help

• Very slow → smaller Search Range, larger Block Step, Estimation → Basic.
• Softness → lower Sigma, smaller Group Size.
• Noise remains → higher Sigma, Estimation → Final, a second slot.
• Flicker between frames → BM3D is spatial; for temporal stability use BM4D or ATAdenoise.

Next — FAQ or the next algorithm: BM4D.