Soft Unload Models Data [LP]

The pack's README describes its unloader family in three tiers: a Hard unloader that fully unloads every model from memory, a Soft one that does "just soft cleaning of memory from garbage" without necessarily evicting model weights, and a dedicated Soft Full Clean RAM and VRAM node for a full unload-and-clean pass at the very end of a queue run. Going by name and behavior, this node - "Soft Unload Models Data" - is that middle tier: a lighter memory nudge you can drop mid-graph, distinct from its SoftFullCleanRAMAndVRAM-LP sibling, which is explicitly the end-of-queue heavy hitter.

When you'd want the lighter touch instead of a full clean

If you're chasing a real out-of-memory error at the very end of a run, you want the full unload-and-clean node placed after your last step. But if you're mid-workflow - between two heavy stages, where you want to nudge memory pressure down a little without paying the cost of a complete model reload on the very next step - this is the better fit. It's a garbage-collection-style sweep, not a scorched-earth unload.

The inputs and outputs that matter

• source - wildcard-typed (*), the value you chain through it, same passthrough pattern as its sibling node.
• Output - the same value, passed straight through.

Notably absent compared to SoftFullCleanRAMAndVRAM-LP: there's no free_execution_cache toggle here. This node's job is narrower - it does the memory sweep and nothing more, it doesn't also clear ComfyUI's node-output cache.

This node is also an output node, meaning it always executes even without anything wired to its output, same as its sibling.

How to install it

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

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

Restart ComfyUI. No models, no extra dependencies.

Common issues & troubleshooting

Memory usage doesn't visibly drop after adding it. This is a known, recurring complaint across the whole category of "unload models" nodes in the ComfyUI community - people reach for a node like this and find RAM or VRAM usage stays flat. It's not something specific to this node that's been confirmed one way or another here, but if you hit it: ComfyUI's own --cache-none startup flag, or the built-in "Unload Models and Execution Cache" action in ComfyUI's settings/hotkey menu, are the more forceful fallback options people report actually working when a soft node-based unloader doesn't move the needle.

You dropped this at the end of your graph expecting a full clean. That's not quite its job - for the end-of-queue full unload plus cache clear, use its sibling SoftFullCleanRAMAndVRAM-LP instead, which is explicitly built for that spot.

You're chaining several unload nodes hoping to stack their effect. There's no evidence that's how this pack's unloaders are meant to be used - each does its own thing at its own scope. If you're not sure which one actually helps your case, start with the lighter mid-graph node here, and only escalate to the full clean at the end if you're still hitting memory pressure.