Passive
The reference and the problem. Everything is scaled by a fixed amount and added together, which is what a standard fold-down plug-in does today. Nothing is measured, and nothing is corrected, so correlated sources will arrive hot.
When panning between two speakers that fold down onto each other, the source will voltage sum rather than power sum, creating about 3 dB of unintentional gain.
This is a specific problem when folding 9.1.6 down to 7.1, where sounds are panned in between the front or the side and the wide, or anything that has height.
pan 61% arriving +1.25 dB intended -1.76 dB excess +3.01 dB
Drag across the plot. The gold line is what arrives at the speaker, the grey line is what you intended, and the gap between them is level the mix never asked for. It is widest in the middle of the pan, where both copies are loudest at once, and closes only at the ends, where they stop meeting.
Correct the voltage summing dynamically.
acFold compares sources that fold onto each other and corrects collisions band by band, leaving uncorrelated information alone. This makes for a very transparent fold-down that simple volume and panning adjustments can't accomplish alone.
There are four automatable engines to help you solve different fold-down scenarios.
The reference and the problem. Everything is scaled by a fixed amount and added together, which is what a standard fold-down plug-in does today. Nothing is measured, and nothing is corrected, so correlated sources will arrive hot.
One gain for each output, band by band. It compares how each fold is summing against what it was intended to sum, and takes back the difference. This is the simplest answer to remove the pan bump caused by summing correlated information. However, the gain lands on everything in the corrected band, so uncorrelated information sharing the band is reduced with it. This engine is great for tracks that have mostly common content throughout, like a dialogue stem that has reverb tails on every channel.
Looks at two contributors together and separates what they genuinely share from what only happens to overlap. Only the shared part is corrected, so a sound that merely sits in the same band as a collision comes through untouched. Its limit is in its name: when three or more copies land in the same output, working a pair at a time cannot account for all of them and some excess survives. This engine is great for tracks that may have sounds panning around a space that share frequencies with different statically panned sounds, like an FX stem with a broadband engine hum and gun shots and ricochets.
Corrects each fold path on its own, before anything sums. A sound's direct path is never touched, so anything already in the right place measures exactly as it went in, and a source arriving from three directions at once is an ordinary case rather than a hard one. The most thorough of the four, its policy setting decides whether it trims only the colliding path, which is level-exact, or spreads the trim across that source's paths to keep its image intact. This engine is great for fold-downs that have a lot of information in many channels, like a music stem.
Every engine removes the same bump. What separates them is what else they touch: content in this output that never collided moves 0.00 dB, because one gain covers the whole band.
Lane by lane, the difference is easier to see. Under Path the sources that never folded read exactly 0.0: nothing was done to them at all.