ATAdenoise

Temporal averaging. It takes neighboring frames and blends them with the current one — but only those frames and only where they really are similar.

What it does

For each pixel it looks at the same pixels in neighboring frames (in a window of 3 to 129 frames, 9 by default). If a neighbor's pixel is similar, it is averaged in; if it differs (motion, a new scene), it is left alone. On a static scene this gives a very clean picture: noise is averaged independently in each frame, while detail matches across frames and therefore doesn't suffer.

When to pick it

• A static scene: interviews, push-ins, talking heads, a locked-off camera.
• You need a clean picture without softness — averaging over time blurs less than spatial blurring.
• Motion compensation is enabled (or there is no motion at all).

When not to pick it: heavy motion without compensation (ghosting), or footage where every frame is fundamentally self-contained.

Parameters in plain words

In Simple, the algorithm leaves you Denoise / Chroma / Balance. In Advanced:

• Ya Threshold and Yb Threshold — similarity thresholds for luma. Ya is "how closely pixels must match to be blended for sure"; Yb is "how much they must differ to be kept apart for sure". Put simply: the thresholds decide how many frames go into the average.
• Ua/Ub/Va/Vb Threshold — the same for the two color components (U and V are the color planes). Color is usually blended more gently.
• Size — the size of the averaging window in frames (odd, 3–129, 9 by default). A bigger window means cleaner static shots, but stronger ghosting on motion and a more expensive render.
• Mode — Parallel (faster) or Serial (more precise). Usually Parallel.
• Sigma — how strong the noise must be to count as noise. It helps the filter tell noise from detail.
• Median — take the median over the accepted frames instead of the mean. The median isn't afraid of isolated "outliers" (artifacts in a single frame), but smears a bit more. Try it if other frames' detail shows up in the average.

The balance between the thresholds is fine-tuning: if the picture is still noisy, raise the similarity thresholds; if ghosting appears, lower Yb or reduce Size.

How it works inside

For each point in the frame the plugin compares it with neighboring frames (with compensation enabled — shifted beforehand to follow the motion). A difference below the Ya threshold means the frame is fully accepted into the average; above the Yb threshold it is fully rejected; between the thresholds it is taken with a partial weight. Thresholds are separate for luma (Y) and color (U, V), because color noise and color detail live on different scales. Median uses the same accepted set but takes the middle sample by value rather than the arithmetic mean — which is why a single artifact in a frame doesn't get stretched into the result.

If it didn't help

• Ghosting / trails on motion → Motion Compensation Enable + Analyze Motion; smaller Size; lower Yb/Ua thresholds.
• Softness → turn Median off, lower the thresholds.
• Noise remains on static shots → bigger Size, higher thresholds, raise Sigma.
• Flicker (noise "breathes" from frame to frame) → bigger Size; this is exactly the case ATAdenoise is picked for.

Next — FAQ or the next algorithm: NLMeans.