VELVET VICE MiniMax H3 — Final Cleanup

Every heavy workflow has a quiet problem nobody enjoys debugging: the previous render leaves the giant models parked in VRAM, so the next render starts with the machine already half-full. The VELVET VICE MiniMax H3 Final Cleanup node is the designated closer - it waits until the final video is actually saved, then reports what the render session did to your memory and releases the tracked models so the next run starts clean.

Its job is sequenced by its inputs rather than by magic. The required final_video input is a VHS_FILENAMES value coming from the video-save/combine node at the tail of the workflow. That socket isn't there because the cleanup node needs to read video data - it's there for ordering. Nothing reaches the cleanup node until the video node has run and produced its filename result, which guarantees the cleanup can't fire early and unload models while they're still being used to encode.

When it finally runs, it does two things:

1. Stops the render memory monitor that the System Check started earlier and summarizes what it saw - peak pressure, warnings, the session's shape.
2. If unload_render_models is on (default), releases the ComfyUI models this render tracked, freeing VRAM/RAM for whatever comes next.

The single output, cleanup_report (STRING), is the human-readable account of all of that - how many models were unloaded, whether the monitor was active, the memory summary. It's an output node, so it also surfaces in the UI, and the same report lands in the console with a [VELVET VICE] FINAL CLEANUP prefix if you're the type who watches logs.

Install

It's part of the one-pack install - no separate setup:

cd ComfyUI/custom_nodes
git clone https://github.com/Velvet-Vice/velvet-vice-minimax-h3

Restart ComfyUI (or Manager, search "velvet-vice-minimax-h3"). No pip dependencies and it owns no model files of its own.

Where it fits and where it bites

In the single-output reference workflow this node sits at the very end, after the video encoder, and it's the difference between a long session that degrades render-by-render and one that stays predictable. The failure mode to know: if you bypass it, or wire it somewhere that lets the graph finish without it, the tracked render models simply stay resident - ComfyUI core does its own unloading eventually, but the pack's memory monitor never gets its summary and the "clean slate" behavior doesn't happen. Also note the node reports on the render session; if you interrupted a run hard (or killed ComfyUI), there's no session left to summarize, and a fresh restart resets everything anyway. It's a tidy-up node, not a savior - if your next render still OOMs after a clean run, the answer is a smaller resolution or fp8, not another cleanup pass.