Eric Qwen-Edit Unload

Eric Qwen-Edit Unload is a cleanup node, and it's exactly as glamorous as that sounds: connect it after your last edit node, and it frees the VRAM the Qwen-Edit pipeline is holding. That's the whole job, and it's a real job - the edit model is a 20B bf16 pipeline that, with keep_in_vram on (the loader default), stays parked in VRAM between runs. If you're bouncing between this pack and another model in the same session, that resident pipeline is the reason your other generation suddenly OOMs.

It's the diffusers-pipeline equivalent of the "unload model" hygiene that ComfyUI power users bake into workflows when they're juggling multiple heavy models in one graph. The pipeline is a big opaque object, not a stack of ComfyUI MODEL/VAE sockets, so the normal ComfyUI way of clearing it (let the graph re-run and overwrite) doesn't apply - you need this node (or a ComfyUI restart) to evict it.

The two ways to wire it

The node takes both inputs as optional, because there are two sensible trigger patterns:

• pipeline - connect your QWEN_EDIT_PIPELINE here to tell it what to unload. This is the straightforward one.
• images - connect an image output (say, from your final Save Image or Preview) as a trigger. ComfyUI only executes nodes whose inputs changed, so wiring the last image in your branch through here makes the unload fire after the generation that produced it, not before. This is the pattern to use when you want to guarantee the unload happens last, in an order you control.

You can wire both, and honestly the images passthrough is the more useful of the two - it doubles as an execution-ordering device. The output is a single status string, which you can pipe into a Text node or just leave floating; its job is to tell you the unload actually ran.

The usual gotcha

Because both inputs are optional and the node is an output node (it runs even with nothing meaningful connected), the failure mode is the opposite of most nodes: it works, but when is up to the graph. If you hook the pipeline into a branch that re-runs on every queue, you can find yourself unloading the pipeline before your edit node in a later execution - the classic stale-order bug the node-plumbing crowd warns about. The fix is to always gate it on the images trigger from your final output, so it's pinned to the end of the chain.

Also worth knowing: this unloads the edit pipeline only. The Eric Qwen-Image Unload node does the same for the text-to-image pipeline, and they're separate - QWEN_EDIT_PIPELINE and QWEN_IMAGE_PIPELINE are different types in this pack, and the README says so explicitly. If you loaded both (you can, with enough VRAM), you need both unload nodes.

Install is the shared pack route - ComfyUI Manager search "Eric Qwen Edit", or git clone https://github.com/EricRollei/Eric_Qwen_Edit_Experiments into custom_nodes/ and restart. Nothing else to download: this node has no model of its own, it just says goodbye to the one your loader brought in.