sections[].noise

Per-section additive Gaussian noise. Optional — omit the block entirely for a clean section.

The block carries two independent layers:

• the capture layer (the top-level keys), describing one playback of the disc;
• the shared layer (shared:), describing the pressing.

    noise:
      noise_db: 43.0
      noise_spread_db: 4.0
      noise_seed: 1001
      spectrum:                # optional
        - [0.0, 0.0]
        - [3.5, 4.0]
        - [6.0, 10.0]
      spread_spectrum:         # optional
        - [5.0, -20.0]
        - [6.0, 0.0]
      shared:                  # optional
        noise_db: 44.0
        noise_seed: 77
        spectrum: [[0.0, 0.0], [5.0, -3.0]]
        level_term:            # optional
          noise_db: 36.0
          level_db: [[0, 0.0], [420, -12.0], [700, -12.0]]
          spectrum: [[5.375, -40.0], [5.625, 0.0]]

Keys

Key	Type	Required	Range	Description
noise_db	float	Conditional	20.0–61.0	Noise floor in dB; sets the Black PSNR target. Required if noise_spread_db is present
noise_spread_db	float	No	0.0 – (noise_db − 20.0)	How many dB noisier white is than black. White SNR = noise_db − noise_spread_db
noise_seed	int	No	any int64	Fixed RNG seed for reproducible noise. Omit for a run-specific seed
spectrum	list of [MHz, dB]	No	see spectrum	Shape of the capture noise spectrum. Requires noise_db. Omit for white noise
spread_spectrum	list of [MHz, dB]	No	see spectrum	Shape of the level-dependent (noise_spread_db) term alone; spectrum then shapes the floor only. Requires noise_db. Omit for the level term to follow spectrum. See spread_spectrum
shared	map	No	see below	The shared (pressing) noise layer

shared

Key	Type	Required	Range	Description
noise_db	float	Yes	20.0–61.0	Pressing noise level in dB
noise_seed	int	Yes	any int64	Seed of the pressing. Captures of one pressing use the same value
spectrum	list of [MHz, dB]	No	see spectrum	Shape of the pressing noise spectrum. Omit for white noise
level_term	map	No	see below	A second pressing component whose level follows the picture's luma. See level_term

shared takes no noise_spread_db. Its constant component is the same whatever the picture; its level dependence, when there is one, comes from level_term, which follows the pressing's own picture and not any capture's signal, so the layer stays identical in every capture.

shared.level_term

Key	Type	Required	Range	Description
noise_db	float	Yes	20.0–61.0	Level of the term where level_db reads 0 dB
level_db	list of [mV, dB]	Yes	see level_term	Gain of the term against the picture's luma level
spectrum	list of [MHz, dB]	No	see spectrum	Shape of the term alone. Omit for white noise

No other key is accepted in any of these blocks.

Behaviour

Configuration	Result
Block absent	No noise
noise_db only	Flat floor; White SNR = Black PSNR = noise_db
Both keys	Two-component model; brighter picture content is noisier
noise_spread_db without noise_db	Parse error
noise_seed only	No noise — the block has no effect
shared only	Pressing noise only; no capture layer
spectrum or spread_spectrum without noise_db	Parse error
spread_spectrum with noise_spread_db 0	Warning: there is no level-dependent term to shape
shared without noise_db or with noise_spread_db	Parse error
level_term without noise_db or level_db	Parse error
Both layers	Their noise values are summed per sample, so the variances add

Noise is injected on the internal millivolt representation before quantisation, so it lands on every synthesised region — VBI lines, blanking and active picture — and survives into the final sample codes. noise_db is the level of each output signal:

output.signal_type	Where the noise goes	Measured level
composite	One noise value per sample, added once (to Y before Y + C is formed)	The composite signal is at noise_db
yc	The same value added to both Y and C (correlated Y/C noise)	The luma and the chroma file are each at noise_db

Y is then clamped to the legal composite range; C is centred on zero, so saturated chroma is not clipped here, and the output stage clamps the combined signal.

spectrum

A table of [frequency_mhz, relative_db] breakpoints describing the shape of the noise power spectrum.

Rule	Value
Points	At least one
frequency_mhz	0.0 to the 4fsc Nyquist frequency, 2 × fsc, strictly ascending: 8.867 MHz (PAL), 7.159 MHz (NTSC), 7.151 MHz (PAL-M). Noise is shaped before the output stage resamples a RAW encoding, so the limit is the same for every sample_encoding_preset
relative_db	−60.0 to +40.0

Between breakpoints the level is a straight line in dB against MHz; below the first point and above the last, the end values are held.

Only the shape matters. The noise is filtered and then normalised so its total power is unchanged, which means noise_db still sets the overall level: a table that tilts the noise towards high frequencies moves power out of the low band into the high band, it does not add any. A one-point table is therefore the same as no table.

The shaping filter is a 255-tap linear-phase FIR designed once per table. Its frequency resolution is about 70 kHz at PAL 4fsc (56 kHz at NTSC 4fsc), and each breakpoint is smoothed over roughly four times that.

On the capture layer without spread_spectrum, noise_spread_db still applies sample by sample: the level-dependent term scales the shaped noise, so bright content carries louder noise of the same shape.

spread_spectrum

A second table, with the same rules as spectrum, that gives the capture layer's level-dependent term a shape of its own. Real captures often grow noisier with level in some bands and not in others; one noise_spread_db with a single shape cannot follow that.

With spread_spectrum present, the capture layer is the sum of two independent terms:

Term	Level	Shape
Floor	σf, from noise_db	spectrum (white when omitted)
Level-dependent	k · Y, from noise_spread_db and the sample's luma level	spread_spectrum

The total power is unchanged — σ² = σ_f² + (k · Y)², so Black PSNR and White SNR are still noise_db and noise_db − noise_spread_db — but the extra power that bright content carries lands only where spread_spectrum puts it. For example, [[5.0, -20.0], [6.0, 0.0]] keeps the noise below 5 MHz almost level-independent and puts the growth with level above 6 MHz.

The shared layer takes no spread_spectrum; its level-dependent part is level_term, which carries its own spectrum.

shared.level_term

On real captures the pressing noise does not keep a constant level: in some bands it falls as the picture gets brighter. A capture-style σ² = σ_f² + (k · Y)² can only grow with level, so the shared layer takes a separate term instead.

With level_term present, the shared layer is the sum of two independent components:

Component	Level	Shape
Constant	shared.noise_db	shared.spectrum (white when omitted)
Level term	level_term.noise_db + the level_db gain at the sample's luma level	level_term.spectrum (white when omitted)

level_db is a table of [level_mv, relative_db] breakpoints:

Rule	Value
Points	At least one
level_mv	0.0 to 800.0, strictly ascending. The luma level in mV above blanking, on the same scale for every standard: white is 700 mV on PAL and SECAM, 714.3 mV on NTSC and PAL-M
relative_db	−60.0 to +20.0: the term's amplitude relative to level_term.noise_db

Between breakpoints the gain is a straight line in dB against mV; below the first point and above the last, the end values are held, so sync and blanking take the first point's gain. A one-point table is a constant second component.

The level is read per sample from the pressing's own picture: the luma as generated, before the channel stage's FM beat, edge smear and timing offset, and before any capture's noise. Every capture of the pressing therefore draws the same term on the same picture, whatever its channel or capture settings. The whole shared layer, term included, is then delayed by channel.timing_offset_samples, so it stays on the picture's samples.

The term has its own seed stream, derived from shared.noise_seed and the disc position, so adding, removing or editing it leaves the constant component's samples unchanged.

For example, this is pressing noise that is level-independent below 5.5 MHz and, above it, 12 dB quieter on mid-grey than on black:

        shared:
          noise_db: 46.0
          noise_seed: 7101
          spectrum: [[5.375, 0.0], [5.625, -40.0]]   # constant part below 5.5 MHz
          level_term:
            noise_db: 36.0                           # above 5.5 MHz, at black
            level_db: [[0, 0.0], [420, -12.0]]
            spectrum: [[5.375, -40.0], [5.625, 0.0]]

The bounds

20.0 dB lower bound — at 20 dB the noise standard deviation is around 10 IRE, roughly 70 mV on PAL, approaching the amplitude of the sync pulse itself. Sync separator reliability degrades below this.

61.0 dB upper bound — above 61 dB the injected noise is smaller than the 10-bit quantisation floor (about 0.085 IRE) and would have no measurable effect.

noise_db − noise_spread_db ≥ 20.0 is enforced for the same reason: the white component cannot be pushed below the floor either.

noise_seed

With a seed, the noise is identical on every run of the project. Without one, each run gets a fresh seed.

Fixed seeds are what make multi-source stacking sets meaningful — see Impairments.

Shared noise and disc position

The capture layer is keyed by section and output position: two projects with the same seed match only when their timelines match.

The shared layer is keyed by disc position only — the CAV picture number when the section carries a picture_number code, otherwise the output position plus the project's first-picture offset. Section boundaries and output position play no part. Two captures of one pressing therefore carry identical pressing noise on the same picture even when one starts later, skips, or replays a range of pictures. Consecutive pictures always get different pressing noise, including the repeated pictures of a held still.

Its seed is mixed with a constant specific to the shared layer, so a shared.noise_seed equal to a capture noise_seed still gives independent noise.

The pressing noise is read off the disc with the picture, so the project's channel.timing_offset_samples delays it exactly as it delays the picture: a capture whose offset differs from the rest of its pressing still carries the pressing noise on the same picture samples, and it cancels in the difference of two aligned captures. The capture layer stays on the output sample grid, since nothing else shares it.

On a CAV disc, a section with shared but without a picture_number code (a lead-in or lead-out, for example) gets a warning: its pressing noise falls back to output position and will not line up across captures that start at different points.

White SNR measurability

The White SNR figure can only be measured from the picture if a VITS carries a white reference on the measurement line: PAL frame line 19, NTSC frame line 20.

If noise_spread_db > 0 and no VITS targets that line, a warning is emitted. The signal is generated correctly regardless; the warning is about whether an analysis tool will be able to verify the target you asked for.