Clear VRAM (Pass Through)

This one does exactly what its name says and nothing more: it unloads every model currently sitting in VRAM, then hands whatever you plugged into it straight through unchanged. It's a manual override button for a problem ComfyUI's automatic model manager doesn't always solve on its own.

Why you'd want this

ComfyUI is generally good about swapping models in and out of VRAM as your graph needs them, but multi-stage video pipelines can outrun that automatic behavior. The node's own description spells out the exact scenario it was built for: "Insert after VAE Decode to free WanVideo/VAE VRAM before VideoCombine." A Wan video workflow typically does a heavy sampling pass, a VAE decode, and then hands frames off to an encoding node like VideoCombine - and if the VAE and diffusion model are still parked in VRAM when encoding starts, you're paying for memory you don't need anymore. Rather than trust the scheduler to figure that out, you drop this node in the wire and force the issue.

How it works

It's a wildcard passthrough. any_input accepts literally anything - IMAGE, LATENT, STRING, whatever you're routing through your graph - and any_output hands the same thing back untouched. The type is *, ComfyUI's "accept anything" convention, so you can splice this node into any wire without a type mismatch. The actual work happens as a side effect: on execution, it forces every model out of VRAM, then lets your data continue downstream.

Inputs and outputs

There's exactly one input (any_input, required, wildcard) and one output (any_output, wildcard). No thresholds, no toggles, no partial flush - it's binary. You either put this node here and it clears everything, or you don't.

Installing it

Through ComfyUI Manager, try searching the pack's title - it's a small, personal repo, so don't be surprised if it's not indexed there. The reliable path, straight from the README, is manual:

cd ComfyUI/custom_nodes
git clone https://github.com/huyl3-cpu/comfyui-sortlist.git

Restart ComfyUI afterward. The README documents this for Google Colab (/content/ComfyUI/custom_nodes), but the same command works anywhere - just point it at your own custom_nodes folder. No models, no extra Python dependencies are mentioned anywhere in the README.

Common issues and troubleshooting

Everything reloads from disk after this node runs, every time. Because it force-unloads all models, not just the ones feeding into it, whatever comes after has to reload from scratch. That's a real time cost. Use it where the author intended - at a genuine bottleneck like the VAE-decode-to-VideoCombine handoff - not sprinkled through your graph out of general caution. Three or four of these in one workflow will make your run noticeably slower for no benefit.

It doesn't show up as an auto-suggested connection. Wildcard-typed nodes like this one don't always play nicely with ComfyUI's drag-to-connect autocomplete, which is driven by matching types. You may need to wire it in manually rather than expect it to pop up as a suggestion when you drag off an output socket.

You can't do a partial flush. If you want to drop just the VAE while keeping a ControlNet or a text encoder loaded, this isn't the tool - there's no selector, it's all VRAM or nothing. For anything more surgical you'd need a node that targets specific models, which this pack doesn't provide.

It's not a substitute for understanding your pipeline's actual bottleneck. If your workflow is still OOMing with this node in place, the issue usually isn't "ComfyUI forgot to free VRAM" - it's that your peak usage (model + VAE + batch size all loaded simultaneously) genuinely exceeds your card's memory, and no flush node fixes that. Lower your resolution, batch size, or reach for a quantized model instead.