[filters] use updated versions of filters + add Luma Key

[sandboxcoder]

• add missing filters: Luma Key, 3-Band Eq, and Upward Compressor
• add version 2 of all known filters that were updated
• Run yarn test:file test-dist/test/regular/api/clipboard.js --match "Copy/paste filters between scene collections" and verify legacy filter is converted into v2 filter

Dependencies

This change is semi-related to this OSN update which filters out disabled/obsolete filters. But that change is not required for this PR.

Regression test cases for QA

This PR updates the following filters: Image Mask/Blend, Color Correction, Color Key, Sharpen, Chroma Key, Noise Suppression, and Luma Key.

This PR adds Upward Compressor (Audio filter), Luma Key Filter (Video), and 3-Band Equalizer (Audio filter)

What changed from version 1 to version 2?

Color Correction - v2 adds proper color space negotiation so the filter works correctly in SDR/WCG pipelines without mangling HDR content, whereas v1 blindly assumed GS_RGBA for everything.

Image Mask/Blend- v2 is a precision and color-pipeline upgrade — higher-resolution opacity and proper sRGB/color space awareness

Color Key Filter- v2 is a substantial rework: opacity is redesigned from an alpha-packed color hack into a proper float parameter, brightness scaling is removed, the contrast/brightness ranges are expanded, and the whole render path gains proper color space awareness.

Sharpness - only adds SRGB capability flag
Chroma Key - the chroma key v2 changes are a mirror of the color key v2 changes — opacity redesigned from an alpha-packed color hack to a proper float, brightness scaling removed, expanded ranges, and full color space pipeline awareness added.

Noise Suppression (Audio) - The change is essentially: v2 promotes RNN noise suppression as the preferred default when available, rather than only falling back to it when Speex isn't present. RNN = Recurrent Neural Network. It refers to a neural network-based noise suppression approach (specifically the https://jmvalin.ca/demo/rnnoise/ library), as opposed to the classic DSP-based Speex noise suppression.

New filters that this PR exposes

Luma Key Filer:
The luma key filter makes pixels transparent based on their brightness (luminance) rather than a specific color. It exposes four parameters:

• Luma Min / Luma Max (0.0–1.0): defines the brightness range to key out. Pixels whose luminance falls within this range become transparent.
• Luma Min Smooth / Luma Max Smooth: softens the edges of the key at the min/max thresholds, so the transparency transition isn't a hard cut.

For example, with default settings (luma_min=0.0, luma_max=1.0, no smoothing), nothing is keyed. If you set luma_min=0.5, pixels darker than 50% brightness become transparent — useful for removing black backgrounds or applying overlays that blend based on brightness
rather than color.

It's commonly used with white/black backgrounds or as a way to composite elements where a color-based chroma key wouldn't work.

Upwards Compressor: The upward compressor is an audio dynamics filter that boosts quiet sounds up toward a threshold, as opposed to a traditional (downward) compressor which turns loud sounds down.

How it works:

1. Detection — measures the signal level using an RMS envelope follower (10ms window)
2. Gain stage — for signals below the threshold, applies positive gain proportional to slope * (threshold - signal_level). Signals above the threshold get 0 gain (left untouched). The slope is 1.0 - ratio, where ratio is 0.0–1.0
3. Knee — a soft knee (0–20 dB wide) smooths the transition around the threshold so the boost doesn't kick in abruptly
4. Ballistics — attack and release times control how fast the gain responds
5. Output gain — a final makeup gain stage

Parameters:

• Ratio (0.0–1.0): how aggressively to boost quiet sounds — higher ratio = more boost
• Threshold (-60–0 dB): the level below which boosting begins
• Attack / Release (1–100ms / 1–1000ms): how fast the gain reacts
• Knee width (0–20 dB): softness of the threshold transition
• Output gain (-32–+32 dB): makeup gain applied after compression

Practical use: leveling out a quiet speaker or microphone so soft parts of speech become more audible without turning up the overall volume — the opposite problem from what a downward compressor solves. It shares the same underlying code as the expander filter, just
with is_upwcomp = true which flips the gain direction.

3 Band Equalizer: It's a simple 3-band equalizer audio filter. It splits the audio signal into three frequency bands and lets you boost or cut each by ±20 dB:

• Low — below 800 Hz (bass)
• Mid — 800 Hz to 5000 Hz (vocals, presence)
• High — above 5000 Hz (air, sibilance)

How it works under the hood:

The crossover frequencies are hardcoded at 800 Hz and 5000 Hz. For each audio sample it runs two 4-pole lowpass filters in series (one tuned to 800 Hz, one to 5000 Hz) using a simple first-order IIR structure. The three bands are then derived by subtraction:

• low = output of the 800 Hz lowpass
• high = delayed input minus the 5000 Hz lowpass output
• mid = delayed input minus high minus low

Each band is scaled by its gain multiplier and the three are summed back together. The EQ_EPSILON added each step prevents denormal floating-point values from causing performance issues.

It's intentionally simple — no adjustable crossover frequencies, no shelf or peak filters, just three fixed bands with gain controls.

[bundlemon]

BundleMon

Files added (4)

Status	Path	Size	Limits
✅	renderer.(hash).js	+7.86MB	-
✅	vendors~renderer.(hash).js	+4.67MB	-
✅	updater.js	+115.29KB	-
✅	guest-api.js	+40.23KB	-

Total files change +12.67MB

Final result: ✅

View report in BundleMon website ➡️

---

Current branch size history

[sandboxcoder]

Closing this one because it can be released as a bundle update. See bundle PR: #6018