Eric Qwen-Image Unload

This node does one thing: it unloads the Qwen-Image generation pipeline from VRAM and tells you it did. That's it. No image output, no fancy plumbing. And it's genuinely worth wiring into any workflow built with this pack, because the thing you're unloading is a 20-billion-parameter MMDiT that eats roughly 40 GB of VRAM on its own.

The QWEN_IMAGE_PIPELINE in the EricRollei pack is the full Hugging Face diffusers pipeline, kept cached in VRAM between runs (keep_in_vram on the loader defaults to True). That's great for iteration speed - each new generation doesn't reload 40 GB of weights. But when you finish a batch and switch to a different model, or start a ComfyUI job with a heavy VAE, a 20B transformer sitting in memory is how you hit an OOM that makes no sense at first glance.

What it looks like in a workflow

Both of its inputs are optional, which is the key detail. You connect:

• pipeline - the QWEN_IMAGE_PIPELINE you want to free
• images - a passthrough from any image output, so the unload fires after generation completes

The images connection is the important one. ComfyUI executes a node as soon as its inputs are ready, so wiring just the pipeline would try to unload before the generate node runs. Hook the image output into images instead, and the unload triggers in the right order - after the picture exists, when you're done with it.

The node outputs a status string ("Generation pipeline unloaded" or "No generation pipeline was loaded") so you can read what happened in the node widget. It's marked as an output node, so it also serves as a clean endpoint for the chain.

Install & gotchas

Same as every node in this pack: ComfyUI Manager, search "Eric Qwen-Edit", or clone https://github.com/EricRollei/Eric_Qwen_Edit_Experiments.git into custom_nodes/ and restart.

Two practical notes. First, this unloads the generation pipeline specifically - the edit pipeline (QWEN_EDIT_PIPELINE) has its own sibling node, "Eric Qwen-Edit Unload." If you're running both, you need both. Second, the loader re-loads the model on the next run, so this is a one-shot release, not a manager that juggles models for you. People running this pack on cards well under the recommended spec (think 8 GB) report that offloading gets brutal - the pack's own memory thread describes 20× slowdowns when sequential offload kicks in. On hardware like that the honest fix is a smaller model, not an unload node, but on properly-sized VRAM this is the clean way to hand the memory back when a generation session ends.