Description
Noise is the number-one enemy of astrophotography.
An astronomical image is by nature a statistical battle: extracting a faint signal (a handful of photons per second per pixel from a diffuse nebula) from a background drowned in various noise sources -- sensor read noise, thermal noise (dark current), sky background noise, and shot noise itself (Poissonian and irreducible).
The signal-to-noise ratio (SNR) is the metric that determines final quality: the higher it is, the smoother the image and the more faint structures emerge from the background. Underexposure describes the situation where the total integration time is too short to reach a usable SNR, resulting in a grainy master where faint structures (nebula extensions, peripheral spiral arms, galactic IFN) are buried in noise.
The defect is universal and largely predictable: SNR grows with the square root of total integration time, so doubling the apparent quality requires four times as many cumulative frames. Underestimating this need is the most common structural error in beginner -- and even intermediate -- astrophotography.
Visual signature
Global grainy appearance across the image, particularly pronounced in the sky background and in areas of faint signal. Fine structures (galaxy arms, nebula filaments, quasar jets, faint narrowband gradients) are buried in an indistinct noise texture.
After stretching, the background looks rough and "dirty" rather than smooth. On an OSC sensor, chroma noise manifests as randomly colored pixels in saturated hues (red, green, blue) against a background that should be neutral. Faint stars partially merge with the noise and lose their definition.
On monochrome narrowband captures, noise produces a "snowy" appearance in low-signal areas that degrades the effective contrast of structures. Objective measurement: in PixInsight via Statistics or NoiseEvaluation, the SNR measured in a region of moderate signal stays below 20 to 30, whereas a well-exposed image easily exceeds 100 to 200.
Differential diagnosis
Distinct from posterization (hard steps in gradients, a consequence of insufficient quantization, not insufficient integration).
Not to be confused with poor seeing (stars are round but bloated, no particular grainy appearance).
Different from failed calibration (residual hot pixels, walking noise, structured point defects -- not a uniform statistical noise).
Not to be confused with excessive read noise from a poorly set gain (a sensor parameter issue, not an integration problem).
Also check that it is not overly aggressive processing amplifying existing noise (a failed denoise can produce a "patchwork" look different from pure noise).
The diagnostic test: measure SNR in several zones via PixInsight Statistics, compare against total integration time and target surface brightness.
If SNR < sqrt(total_time/unit_time) x expected_unit_SNR, integration is simply too low for the target.
Not to be confused with over-denoising: noise from underexposure is real grain caused by insufficient signal, whereas over-denoising is the "plastic" artifact created by trying to mask that grain in post-processing. One is the cause, the other is the wrong response.
Probable causes
- Insufficient total integration time for the surface brightness of the target
- Particularly faint target (IFN, galaxy tidal tails, polar jets) underestimated during planning
- Unit exposure too short with too low a gain (read noise dominates)
- High-thermal-noise sensor without cooling (DSLR, unregulated sensor)
- High light pollution at the site, degrading SNR by raising the sky background noise floor
- Unsuitable filter passing too much light pollution (no LP or dual-band filter under Bortle 7-9 skies)
- Sensor gain poorly set for the target dynamic range (too low: read noise proportionally higher)
- Bright Moon saturating the sky background and reducing effective contrast
- Narrowband filters with unit exposures that are too short (Ha/OIII at 60 s instead of 300 s)
- Stacking very few frames (5 to 10) where 50 to 100 would be needed
Course of action
- Calculate the target integration time: aim for more than 2 to 3 hours in moderate RGB, more than 5 to 8 hours in narrowband or on a faint target
- For very faint targets (IFN, extragalactic halos, surface magnitude >23), plan across multiple nights for 10 to 20 cumulative hours
- Match gain to the image type: unity gain (HCG on ASI/QHY) for long narrowband; low gain for short bright-target sessions
- Set unit exposure so sky background noise clearly dominates read noise (test: histogram peak at 1/4 to 1/3 of dynamic range)
- On light-polluted sites, use an LP filter (Optolong L-Pro, Hutech IDAS LPS) or switch to dual-band (L-eXtreme, L-Ultimate, NBZ)
- Cool the sensor to -10 deg C or -20 deg C per spec to minimize thermal noise
- Stack with an appropriate algorithm: sigma clipping rejection on more than 20 frames, local normalization
- For multi-session projects, recalibrate the dark set whenever temperature shifts between seasons
- On an already underexposed master, NoiseXTerminator, NoiseExterminator, and MultiscaleMedianTransform attenuate visible noise without creating the missing signal
- Accept that some targets require several dedicated nights and plan accordingly
The Doc's advice
Noise is pure physics: you double the SNR by quadrupling integration time, full stop.
There is no magic shortcut, no miracle denoise that replaces acquisition hours. NoiseXTerminator is great for smoothing a solid master; it is useless for salvaging a 45-minute session on a magnitude-11 galaxy.
If your image is too noisy, the question to ask is not "what filter should I apply" but "when am I going back to shoot it." And accept an uncomfortable truth: half the beautiful images you see on the forums have 15 to 30 hours of integration behind them.
Think you can see this defect in your image?
Run a diagnosisFrequently asked questions
How much integration time is needed for a given target?
It depends on the surface brightness of the target, the aperture of the instrument, sky quality, and image type. Practical benchmarks: a bright galaxy (M51, M81, M101) needs 3 to 5 hours in RGB for a clean result; a classic emission nebula (M42, NGC 7000) needs 2 to 4 hours in RGB or 4 to 6 hours in HOO; a faint object (IFN, tidal tails, galactic halos) needs 10 to 30 cumulative hours; extreme details (HH jets, faint planetary nebulae) can demand 50 hours or more. The practical rule: if the image is still grainy after stretching, more frames are needed, period.
Should you use long exposures or stack many short ones?
It depends on the ratio of read noise to sky background noise. The rule of thumb: the unit exposure must be long enough for sky background noise to clearly dominate read noise (by a factor of 3 to 10). Under light-polluted skies (Bortle 6 to 9), 60 to 180 s is often sufficient. Under dark skies (Bortle 2 to 4), exposures of 300 to 600 s are needed to swamp the read noise. Stacking many very short frames (<30 s) under dark skies wastes SNR: you accumulate a great deal of read noise without proportional gain. Conversely, very long exposures (>600 s) under light-polluted skies saturate the sky background without useful gain. Find the balance with a simple test: the median background on a unit frame should sit at roughly 1/3 of the sensor's dynamic range.
Can NoiseXTerminator replace a lack of integration?
No, but it can mask the visual consequences. NoiseXTerminator (and its open-source equivalent NoiseExterminator) uses a neural network to smooth noise while preserving structures. On a properly exposed master, it turns a slightly grainy image into a very smooth one -- that is its legitimate use. On a severely underexposed master, it produces a smoothing that looks plastic, and faint structures that were not in the data remain invisible: the actual SNR does not change, only the cosmetic appearance does. The usage rule: NoiseXTerminator as a finishing step, never as a substitute for real integration.
Is chroma noise worse on an OSC than in mono RGB?
Often yes, for two reasons. First, the Bayer matrix effectively divides pixels by channel: for each red or blue pixel, only 1/4 of the photosites contribute; 1/2 for green. Per-channel noise is therefore structurally higher than in mono, where every pixel collects all the photons. Second, debayering interpolates the missing channels, which amplifies local noise through diffusion. Practical compensation: to achieve the same per-channel SNR as mono, plan for 2 to 3 times more integration time on an OSC. This is one reason why the switch to monochrome is often the most impactful upgrade in final quality, ahead of changing the tube or the mount.