ComfyUI Node

Controlled VAE Loader

A VAE loader you can gate — the smallest node in the pack, honestly

By Isi-dev·Created about a year ago·Updated 8 months ago· 5
Controlled VAE Loader
  • trigger
  • VAE
vae_name

Controlled VAE Loader is the most honest wrapper in Isi-dev's pack: it's the stock VAELoader with a trigger input welded on, and not much else. Fire the trigger and it loads the VAE and hands you a VAE output. Leave it unplugged and it returns None - no load, no VRAM spent on a VAE you didn't need in that branch.

The VAE is the smallest of the three big weights in a diffusion workflow, so this is the node with the least to gain - a VAE is rarely the thing that OOMs you. Where it does earn its keep is workflow branching: if you have two model paths in one graph (say a checkpoint that bundles its own VAE and a separate VAE you only want under certain conditions), the trigger lets you skip the extra load cleanly instead of leaving a loader that always executes. The wrapper delegates to VAELoader.load_vae(), so the output is a bog-standard VAE that plugs into the VAE input of a KSampler or the VAE Decode node.

The inputs that matter

Inputs are thin by design: the vae_name file picker and the trigger. Like every other Controlled node in this pack, the pause path returns None - wire that into a VAE Decode and you'll get an error about a None VAE, not a graceful skip. It's a switch, not a maybe.

Where it actually fits

If your actual goal is memory management, the VAE is the wrong place to focus anyway. The model that kills low-VRAM runs is the UNet or the text encoder - that's what the pack's Delete Model node and the Controlled UNet / GGUF loaders are for. And since a VAE only lives in VRAM during the encode/decode steps anyway, ComfyUI's built-in management already handles it fine for most people. This node exists because the pack author wanted every loader gated the same way for consistency. Fair enough - it costs nothing to install with the pack - just don't expect it to solve a VRAM problem on its own.

Install is shared with the rest of the pack: ComfyUI Manager (search "DeleteModelPassthrough") or git clone https://github.com/Isi-dev/ComfyUI_DeleteModelPassthrough into custom_nodes, then restart. No model downloads, no dependencies beyond psutil.

CategoryMemory Management

Inputs (2)

NameTypeDefaultDescription
vae_nameCOMBO1 options: pixel_space
trigger*

Outputs (1)

NameTypeDescription
VAEVAE