ComfyUI Node
RedNode Free VRAM
Unloads models at this exact point in the chain, for running a second big model (a detailer, an upscaler) on a card that cannot hold both. Whatever is wired in comes straight out; the wire is what fixes the timing. Nothing needs reloading by hand.
RedNode Free VRAM
- value
- keep
- keep_2
- value
◄unload_modelstrue►
◄empty_cachetrue►
◄always_runfalse►
◄enabledtrue►
CategoryRedNode/Control
Inputs (7)
| Name | Type | Default | Description |
|---|---|---|---|
| value | * | anything — it comes out untouched. Wiring it through is what fixes WHEN the memory is freed. | |
| unload_models | BOOLEAN | true | unload every model ComfyUI is holding. They reload on demand, so the cost is load time, not the run. |
| empty_cache | BOOLEAN | true | hand the freed blocks back to the driver. Leave on: unloading alone often leaves VRAM looking just as full. |
| always_run | BOOLEAN | false | free even when nothing upstream changed. Costs you the cache: everything downstream recomputes every queue. |
| enabled | BOOLEAN | true | off means pass straight through and free nothing. Here so a Control Panel or Combo Control can switch it, which bypassing the node cannot do. |
| keepopt | * | a model to leave resident — wire the one you are about to use again. LoRA-patched clones count as the same model. Models ComfyUI does not manage (SAM3, most upscalers) are never unloaded by this node in the first place. | |
| keep_2opt | * | a second model to leave resident, for a checkpoint and its CLIP or VAE. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| value | * | — |