Hunyuan 3 Unload

ComfyUI's own "Free" button is fine for most models. For an 80B one that fills 45–80GB of VRAM, you want a deliberate unload - and that's what Hunyuan 3 Unload (class HunyuanImage3Unload) is. It's the pack's baseline memory release: it clears the cached Hunyuan model out of VRAM so the next model in your pipeline - or the next run entirely - starts from a clean slate.

Think of it as the middle option in the pack's unload lineup. Soft Unload parks the model in CPU RAM for a fast ~5 second restore. Force Unload (Nuclear) blasts everything including the kitchen sink. This one sits between them: it actually frees the Hunyuan weights, doesn't touch anything else, and gives you a straight answer about what it did.

How it works

Two toggles, one of which does most of the work:

• force_clear (default True) - the actual unload. Clears the cached Hunyuan model from VRAM.
• clear_for_downstream (default False) - the interesting one. When True, it aggressively clears VRAM for downstream models (Flux, SAM2, that kind of thing). The README frames it for successive runs where Hunyuan has to reload anyway: unload Hunyuan hard, free the card, let Flux do its thing, and don't pretend you're going to need Hunyuan cached again this session.

There's also an optional trigger (any type) so you can fire it off the end of a generation chain rather than manually.

The inputs and outputs

Inputs are just the two toggles above plus trigger. Outputs:

• cleared (BOOLEAN) - whether the unload actually happened.
• vram_status (STRING) - a text readout of VRAM after cleanup.
• unload_signal (wildcard) - a passthrough you can feed into a loader's unload_signal input to sequence "unload now, then load" cleanly.

How to install it

It ships in the Comfy_HunyuanImage3 pack:

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

Restart ComfyUI, hard-refresh the browser, and it appears under HunyuanImage3. Or install via ComfyUI Manager (search HunyuanImage).

Common issues & troubleshooting

The README's utility-node table is worth internalizing, because reaching for the wrong unloader is a real trap. This one is for between workflows and manual control - standard cleanup, slow reload is fine. If you're switching between Hunyuan and another model repeatedly and want fast restores, use Soft Unload instead (model stays in RAM). If VRAM is genuinely stuck or polluted from another tab, use Force Unload with the nuclear options. If you need Hunyuan to stay loaded while clearing Flux/SAM2, the pack's Clear Downstream Models node is the one that keeps Hunyuan and drops everything else.

One more honest note: on 24GB cards running NF4 with heavy CPU offload, a plain unload then a reload can take minutes - that's not a bug, it's the model being ~45GB. On that hardware the Soft Unload path is usually the better daily driver.