Documentation · Premiere Pro
FFTdnoiz
Balance. It cuts the frame into blocks and cleans each block by frequency — a compromise between speed and quality.
What it does
The frame is divided into small blocks (8×8 by default). Inside a block the transform separates the "smooth" part of the picture (signal) from the "fine" part (noise). Noise is suppressed by the given strength, the signal remains. Blocks are stitched back together with overlap so the seams are invisible.
When to pick it
- You need a balance: HQDN3D is too weak, while NLMeans/BM3D are too slow.
- Medium noise on ordinary footage.
- A second slot after a fast pass.
When not to pick it: reference quality (BM3D/BM4D are better), or very large detail that the block-based nature of the algorithm can damage at extreme settings.
Parameters in plain words
- Sigma — how strong the noise must be to count as noise. The main slider: higher — more aggressive cleanup. Start at 1 and raise it until the noise goes away.
- Amount — the share of the possible cleanup (0.01–1). A gentler version of "sigma": 0.5 means half.
- Block Size — the block size (4–16, 8 by default). A bigger block means stronger noise suppression, but a higher risk of block artifacts at strong settings.
- Overlap — block overlap (0–0.75, 0.5 by default). The larger it is, the less visible the seams; it runs slower.
- Method — the suppression method:
Wiener Fast— adaptive suppression, fast. The working mode.Hard Fast— hard threshold, fast. Can give a "watery" picture at strong values.Wiener Quality/Hard Quality— the same, but with more careful weighting of the overlaps. Slower, cleaner; for the final render.
In Simple: Denoise moves Sigma, Balance moves Block Size and Overlap (further left is faster, further right is better quality).
How it works inside
Each block is represented as a sum of "smooth" low-frequency components and "fine" high-frequency ones. Noise lives mostly in the high frequencies, but so does detail — except detail is correlated between blocks and noise is not. The Wiener variant attenuates each frequency in proportion to the estimated "signal vs noise" ratio; the Hard variant simply cuts frequencies below the threshold. Overlap with weighting is needed so that suppression in one block doesn't leave a visible grid at the seams.
If it didn't help
- Block artifacts (a grid) → more Overlap, smaller Block Size, Method to a Quality variant.
- Still noisy → higher Sigma, a second slot (BM3D).
- Softness → lower Sigma, lower Amount.
- Slow → Method to Fast, less Overlap, Balance to the left.
Next — FAQ or the next algorithm: BM3D.