Soft Full Clean RAM and VRAM [LP]

The pack's own README describes this one directly: it's for "the case when it is necessary to unload all models from memory and clean memory from garbage after the workflow is completed" - i.e., a terminal cleanup step you drop at the very end of a graph, once the queued task is actually finished, not a mid-workflow memory nudge.

Where it goes in your graph

This is an output node, which in ComfyUI terms means it always executes, whether or not anything is wired to its output socket. That's deliberate - you're meant to chain it after your last real step (typically your Save Image node) so it fires once, at the tail end of the run, after everything that actually needed the models is already done.

The inputs and outputs that matter

• source - wildcard-typed (*), the wire you chain in from whatever should finish first. This is what forces execution order: ComfyUI won't run this node until whatever feeds source has completed.
• free_execution_cache - a boolean, default true. Beyond unloading models, this also clears ComfyUI's cached node outputs, so a repeat run doesn't hold onto stale tensors from the last one.
• Output - passthrough of the same value you fed into source, in case you need to chain anything after it (usually you won't; it's typically the last node in the graph).

How to install it

ComfyUI Manager: search "ComfyUI Level Pixel" or ComfyUI-LevelPixel, install, restart - auto-updates via Manager's "Update ALL." Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/LevelPixel/ComfyUI-LevelPixel.git

Restart ComfyUI. No models, no external dependencies.

Common issues & troubleshooting

You add it and your RAM/VRAM usage doesn't visibly drop. This is a real, reported problem across the whole category of "unload models" nodes, not something specific to this one that I can confirm either way - people reaching for third-party unload nodes and finding memory usage stays flat is a recurring complaint in the ComfyUI community. If that happens to you here, ComfyUI ships its own more forceful options worth trying as a fallback: launching with the --cache-none startup flag (which stops ComfyUI from caching node outputs in the first place), or using the built-in "Unload Models and Execution Cache" action from ComfyUI's settings/hotkey menu, which several users report as the thing that actually worked when a node-based unloader didn't.

You expected it to run mid-workflow, between two heavy steps, and it fired at the wrong time. That's not what this node is for - it's the end-of-queue variant. For a lighter, mid-graph memory nudge without a full unload, the pack's SoftModelUnloader-LP sibling is the better fit; save this one for the tail of the graph.

You're chaining something after the output and it's not behaving as expected. Remember the output is a straight passthrough of source - this node doesn't transform the value at all, it only performs the cleanup as a side effect while passing your original data through unchanged.