Performance guidelines

Video effects, audio effects and generators can execute many times per second. Treat processor functions as hot paths and move expensive analysis out of the real-time render loop whenever possible.

Hot-path guidelines

• Reuse immutable lookup tables, kernels and plans instead of rebuilding them per frame.
• Cache expensive data by parameter signature when reuse is safe.
• Prefer vectorized NumPy operations over Python per-pixel or per-sample loops.
• Avoid filesystem scans, network requests and model loading inside frame/audio processors.
• Return early for provably neutral parameter states.
• Avoid unnecessary copies of full-resolution frames or long audio arrays.
• Keep generator output exactly at the requested size.

Use host metadata correctly

1. Set time_dependent accurately.
2. Add time_dependent_when when only some parameter states are animated.
3. Add bypass_when for states that are exactly neutral.
4. Use a timeline processor for expensive analysis that does not belong in the real-time path.

Tip: For model-based tools, load the model once per processor run or cache it at module scope when the underlying library is safe to reuse.