Video noise is the random variation in brightness and color that appears in moving images as grain, speckles or a shimmering texture which changes from frame to frame and carries no picture information. In video quality assessment it counts as one of the most visible impairments, because the eye is highly sensitive to temporal fluctuation in flat areas such as skies, walls and skin.
Noise originates mainly in the image sensor and its readout electronics, but it is amplified by gain, shaped by processing and then interacts with compression. Understanding where it comes from explains why a clean image in Video depends as much on light and exposure as on the camera or the encoder.
Where video noise comes from
Video noise arises from several physical and electronic sources that add unwanted signal to each pixel. The most fundamental is photon shot noise: light arrives as discrete photons, so the count collected by a pixel fluctuates statistically, with a standard deviation equal to the square root of the signal. Dim scenes therefore have a lower signal-to-noise ratio no matter how good the camera is.
Electronics add their own contribution. Read noise comes from the amplifier and analog-to-digital converter, dark current grows with sensor temperature and exposure time, and fixed pattern noise results from small differences between individual pixels or columns. The following table summarizes the main sources and their visual signatures.
| Source | Cause | Typical appearance | Worst condition |
|---|---|---|---|
| Photon shot noise | Statistical arrival of photons | Fine random grain | Low light, deep shadows |
| Read noise | Amplifier and converter electronics | Grain in the darkest tones | High gain, underexposure |
| Dark current | Thermal electrons in the sensor | Hot pixels, warm speckle | Long exposure, hot camera |
| Fixed pattern noise | Pixel and column mismatch | Static lines or banding | Pushed shadows in grading |
| Quantization noise | Too few bits per sample | Stepped gradients, dither | 8-bit capture, heavy grading |
The distinction between random and fixed pattern noise matters in practice. Random noise changes every frame and reads as a living texture, while fixed pattern noise stays in place and becomes obvious when the camera pans, because the pattern appears to float over the moving scene.
Luminance noise, chroma noise and temporal behavior
Noise is usually described by what it affects and how it moves. Luminance noise varies brightness and looks similar to film grain, whereas chroma noise produces colored blotches of magenta and green that look unnatural and are generally judged more objectionable. Because human vision resolves color detail less finely than brightness, chroma noise can be filtered aggressively with little visible loss of sharpness.
The temporal dimension separates video noise from still-image noise. A single frame may look acceptable, yet in playback the frame-to-frame flicker of noise draws attention, especially at 24 or 25 frames per second on large displays. The table below compares the common categories.
| Category | What varies | Perceived effect | Typical remedy |
|---|---|---|---|
| Luminance noise | Brightness per pixel | Grain-like texture | Moderate spatial and temporal filtering |
| Chroma noise | Color per pixel | Colored blotches | Strong chroma filtering |
| Temporal noise | Pixel values across frames | Shimmer and crawling | Motion-compensated temporal filtering |
| Spatial pattern noise | Fixed positions in frame | Lines or blotches that float | Calibration and pattern subtraction |
Noise is often confused with neighboring impairments in the same Topic. Mosquito noise and blocking artifacts are created by the encoder, not the sensor, and banding is a quantization problem in smooth gradients. Sensor noise can actually hide banding, which is why some encoders add a controlled amount of grain back after denoising.
How gain, exposure and sensor design affect noise
The visible noise level depends chiefly on how much light reaches each photosite and how strongly the signal is then amplified. Raising gain or ISO does not create shot noise, but it lifts the whole signal, including read noise, closer to the bright end of the recorded range, so underexposed footage brightened in post looks noisier than correctly exposed footage.
Larger photosites collect more photons per frame, so sensor size and pixel pitch influence noise performance. Many cinema cameras also use dual native ISO designs, which switch to a second, higher-gain readout circuit with lower read noise for dim scenes. Each one-stop increase in gain roughly halves the signal-to-noise ratio in the shadows, although the exact figure depends on the sensor.
| Factor | Change | Effect on noise | Trade-off |
|---|---|---|---|
| Gain or ISO | Increase by one stop | Visibly more grain | Brighter image without slower shutter |
| Exposure | Open aperture one stop | Lower relative noise | Shallower depth of field |
| Shutter angle | From 180 to 360 degrees | About one stop cleaner | More motion blur |
| Sensor size | Larger format, same resolution | Lower noise per pixel | Larger lenses and cost |
| Bit depth | From 8-bit to 10-bit | Less quantization noise | Higher data rate |
Noise, compression and measured quality
Noise is expensive to compress because it is random and therefore unpredictable. Codecs such as H.264, HEVC and AV1 rely on similarity between neighboring pixels and successive frames; random grain defeats both spatial and temporal prediction, so a noisy source needs a much higher bitrate to reach the same visual result as a clean one.
When the bitrate is constrained, the encoder discards the noise unevenly. The result is smeared, blotchy texture, crawling patches in dark areas and amplified compression artifacts. For this reason streaming services commonly apply denoising before encoding and sometimes describe the removed grain parametrically so that the decoder can resynthesize it, as with the film grain synthesis tool in AV1.
Objective metrics handle noise imperfectly. PSNR penalizes noise strongly because it measures pixel error, while SSIM and VMAF weigh structure and perceived detail, so a denoised image can score well even when it looks plastic. Reliable evaluation of noise still requires viewing tests on target displays.
Reducing video noise without losing detail
The most effective noise reduction happens at capture: more light, correct exposure and the native gain setting outperform any software filter. In post-production, temporal noise reduction compares several frames and averages static detail, while spatial noise reduction smooths each frame on its own and is more likely to soften texture.
| Situation | Recommended approach | Reason |
|---|---|---|
| Static or slow shots | Temporal filtering with motion estimation | Averages noise while keeping detail |
| Fast motion or handheld | Light spatial filtering | Temporal filtering may cause ghosting |
| Strong color blotches | Chroma-only filtering | Removes noise with little sharpness loss |
| Delivery at low bitrate | Denoise, then add fine grain | Saves bits and masks banding |
- Over-filtering produces waxy skin and loss of fine texture.
- Aggressive temporal filtering smears moving edges and causes ghosting trails.
- Denoising after sharpening locks in enlarged grain, so denoise first.
- Grading before denoising can push shadow noise beyond what filters can repair.
In simple terms
Video noise is the fizzy, sandy texture that appears on footage filmed in poor light. A camera sensor works a little like a bucket catching raindrops of light. When there is plenty of rain, small random differences between buckets do not matter. When the rain is light, those random differences become visible as dancing specks.
Turning up the camera’s brightness setting is like turning up the volume on a quiet radio station: the voice gets louder, but so does the hiss. The best cure is more light, and software cleanup works best when used gently.
A practical example
A documentary crew films an interview in a dim apartment lit by one window. The camera has a base sensitivity of ISO 800 and a second native setting of ISO 3200. Shooting at ISO 800 underexposes the subject by about two stops, so the editor would have to brighten the image later, which would lift shadow noise and chroma blotches in the dark furniture.
The camera operator switches to the second native ISO of 3200, keeps the shutter at 1/50 s and adds a small LED panel bounced off the ceiling. In post, the colorist applies mild temporal noise reduction before grading. The final 1080p delivery encoded at 8 Mb/s shows clean skin tones with no visible blotching or crawling texture, and the encoder spends its bits on detail rather than random grain.