Compressing inactive TOP networks to one-eighth resolution recovers GPU memory headroom in multi-scene VJ rigs

In a VJ rig with many pre-loaded visual scenes, most TOP networks are not rendering to the output at any moment. A Python script can traverse the network tree, find all TOP operators in inactive branches, and set their resolution to 1/8 of their full size. When a scene becomes active, the script restores full resolution. This avoids the stutter associated with VRAM flush/clear operations and keeps the GPU from hitting its memory ceiling. The technique is distinct from turning operators off (which stops cooking but can cause pop-in) and works continuously in the background. Hardware limits remain; the script only buys headroom.

Examples

AAVJ VJ mixer implements a VRAM cleanup script: all cells not currently playing are compressed 8×. When a cell is launched, VRAM briefly spikes as it upscales, then stabilises. The VRAM monitor in the interface shows the freed headroom.

Assessment

Explain the trade-off between (a) disabling inactive TOPs, (b) VRAM flushing, and (c) resolution-compression for managing GPU memory in a live rig. When would each approach cause a visible artifact?