Hunyuan Unload V2

The V2 rewrite in this pack replaced a tangle of memory-management nodes with one philosophy: generate nodes carry a post_action dropdown (keep_loaded / soft_unload_to_cpu / full_unload), and the cache system handles the rest. Hunyuan Unload V2 (class HunyuanUnloadV2) is that same philosophy as a standalone utility - the minimal unload node. One dropdown, no outputs, no confirm gates, no nuclear toggles. If you just want a "clean up now" node that matches the pack's modern cache system, this is the one.

It's worth being clear about what it is not. It's not the nuanced one - the pack still has Hunyuan 3 Unload (with clear_for_downstream for handing off to Flux/SAM2) and Force Unload (Nuclear) (for stuck VRAM and orphaned tensors). Unload V2 is the plain version for when you don't need any of that nuance: you want Hunyuan out of memory, one way or another.

How it works

One required input, an unload_type dropdown with exactly two choices:

• soft_unload (default) - park the model in CPU RAM. Frees VRAM, keeps the weights warm for a fast restore (~10–30s) the next time a loader needs them.
• full_unload - free VRAM and RAM. The model is gone entirely; the next load is a full disk reload.

That's the entire control surface. It's an output node with no outputs - you drop it in the graph, set the mode, and queue. Nothing to wire, nothing to misconfigure.

How to install it

Same pack, same routine:

cd ComfyUI/custom_nodes
git clone https://github.com/EricRollei/Comfy_HunyuanImage3
cd Comfy_HunyuanImage3
pip install -r requirements.txt

Restart ComfyUI and hard-refresh the browser. Or install via ComfyUI Manager (search HunyuanImage).

Common issues & troubleshooting

The choice between the two modes is really the whole "issue" surface, so here's the honest guidance:

• Soft unload is the right default for a working session - successive runs stay fast because the loader restores from RAM. But the parked model occupies 45GB+ of system RAM. On a 64GB box running NF4, that plus OS and ComfyUI overhead gets uncomfortable.
• Full unload is for when you're genuinely done with Hunyuan or RAM is precious. Your next Hunyuan run pays the full reload cost - minutes on a big model, not seconds.

One practical gotcha: if you set soft unload but the next run still reloads from disk, check whether your model was loaded with device_map offloading - soft unload needs offload_mode='disabled' (no meta tensors) to work, same requirement as the dedicated Soft Unload node. In that case full_unload is the mode that reliably does what it says. For anything more surgical - clearing downstream models while keeping Hunyuan, or nuking stuck allocations - the other unload nodes in this pack exist for a reason.