Motion Compensation

Temporal algorithms (ATAdenoise, HQDN3D with its temporal part, BM4D) clean a frame by averaging it with its neighbors. The problem: if there is motion in the shot, neighboring frames do not match the current one — and averaging produces ghosting. Motion compensation solves this: before averaging, each neighboring frame is shifted along with the motion.

What to turn on

The Motion Compensation group in Effect Controls:

• Enable — motion compensation. For temporal algorithms, keep it on; without it you get ghosting on motion.
• Algorithm — how to compute motion: Auto, Block Match (CUDA), Block Match (CPU), NvOF (GPU). Auto picks for your hardware and settings — in 99% of cases it is exactly what you want.
• Precision — precision vs speed: Fast, Balanced, High Quality, Precise, Fine Detail, Max Detail. Balanced is the working mode; Fast is for heavy footage in real time; the upper steps are for the final render.

In Simple mode the whole group collapses to a single Motion Comp list: Off, Fast, Balanced, High Quality, Precise, Fine Detail, Max Detail — that is, Enable and Precision on one row. Off — the same compensation, turned off; the other steps turn it on at the corresponding precision. The row appears only when slot 1 holds an algorithm that actually reads the motion field: HQDN3D, ATAdenoise or BM4D. For spatial algorithms (NLMeans, FFTdnoiz, BM3D) compensation changes nothing, so there is no list there.

The rest — Algorithm (how motion is computed), Temporal Gate, Post-Processing — lives in Advanced.

The Analyze Motion button

By default, motion is computed fresh on every frame — on scrubbing and rendering that costs noticeable time. Analyze Motion computes motion vectors once, in advance:

1. Put the playhead on the clip (or select the clip).
2. Press Analyze Motion in the Motion Compensation group.
3. The plugin walks the clip and writes the vectors to the cache. Progress is shown on the button; the cache is stored in a hidden effect parameter — that is, it is saved with the project.

After that, every temporal algorithm reads the vectors from the cache: cleanup runs on frames that really match, and there is no need to recompute per frame. Pressing the button again recomputes the cache — useful if the clip or the motion settings have changed a lot.

What to know about the cache

• The cache lives in the project: close and reopen — the vectors are still there.
• The cache takes RAM (on the order of hundreds of megabytes for a long 4K clip); if a frame has not been used for a long time, its vectors are unloaded automatically.
• The cache is tied to the frame and to the motion field settings: change the size, frame rate or geometry — and the cache recomputes itself.
• The Clear — N frames row on the button is an offer to reset the cache (for example, before handing off a project, or when you suspect it is stale).

Without Analyze Motion

Everything works without the button too — each render simply computes motion itself. On short clips you will not notice; on long ones and in 4K it spends time on every pass. Analyze Motion is an optimization, not a mandatory requirement.

Temporal Gate

The Temporal Gate subgroup decides when neighboring frames can be trusted:

• Temporal Gate Mode: Off, Motion Only, Cut Only, Motion + Cut (the default mode). Motion + Cut — the filter does not blend frames where motion is strong, and does not blend across cuts.
• Gate Threshold — how strong motion counts as "too strong".
• Cut Threshold — sensitivity to cuts (scene changes).
• Motion Scale — the gate's overall sensitivity.

Practical meaning: the gate keeps temporal cleanup from "dragging" across fast motion and at frame boundaries. If there is no ghosting, you can leave it alone.

Post-Processing

The Post-Processing subgroup edits the motion vectors themselves:

• Vector Smooth — smoothing of the vector field (Smooth Strength, Type: Median / Bilateral / Gaussian, Radius). Helps when the vectors "jitter" and the picture micro-shakes because of it.
• Confidence Mask — the filter does not trust vectors equally everywhere: at object boundaries, in fast motion and where the search is unsure, blending is reduced (Residual Threshold, Residual Strength, Image Edge Strength, Flow Boundary Threshold/Strength, Confidence Refine Radius/Range). If halos appear on contours — look here.

How it works inside

Motion compensation estimates where each block of the frame moved between neighboring frames, and shifts the neighboring frame back — so that it matches the current one. Only then do temporal algorithms average the frames: the difference "current minus shifted neighbor" turns out to be small exactly where there is no motion in the shot and no new scene has started.

Confidence Mask computes that difference and uses it as a measure of trust: where the difference is large — the object is moving fast or the search was wrong — the contribution of the neighboring frame is reduced. Vector Smooth blurs the vector field itself, removing isolated outliers.

Motion estimation is the most expensive part of temporal cleanup, which is why the vectors are cached: computed once, they are reused by every algorithm and every frame.

If it did not help

• Ghosting on motion remains → Balance further left, Temporal Gate to Motion + Cut, Precision higher.
• The picture micro-shakes → Vector Smooth, Confidence Mask.
• All good but slow → Analyze Motion (cache), Precision Fast, Draft Preview.

More details — FAQ.