Load VAE (No Dynamic VRAM)

The VAE is the smallest model in your workflow, which is exactly why ComfyUI's dynamic VRAM machinery is the wrong tool for it. ComfyUI's newer mmap-backed loader can page weights in and out of VRAM on demand - great for a 32B transformer, pointless for a VAE that weighs a few hundred MB. This node is the escape hatch: it loads a VAE the direct way, bypassing the "aimdo"/dynamic VRAM layer entirely, which is what the "(No Dynamic VRAM)" in its display name means.

If you've ever watched a workflow stall at the VAE load step or felt like your VAE was hogging RAM it had no business using, this is the loader you swap in. It's a drop-in replacement for the stock Load VAE.

How it works

The mechanism is simple and that's the point. Instead of handing the file to ComfyUI's mmap-based dynamic loader, the node parses the safetensors directly using the unifiedefficientloader package (the pack's one real dependency), then builds the VAE with comfy.sd.VAE(sd=..., metadata=...) and lets ComfyUI's own architecture auto-detection do the rest.

Note the source hardcodes the static path - disable_dynamic=True - because comfy.sd.VAE() doesn't accept a dynamic-loading parameter anyway. A VAE is small enough to just sit in VRAM; paging it from disk is pure overhead with no upside.

The inputs that matter

• vae_name - pick your VAE from the vae folder, same as the stock loader.
• low_memory - the one toggle, and the tooltip is the author's own words: "Use fast and efficient low impact loading of model. Set to False to use comfy's default loading." Default is False, so if you want the fast, low-impact path, flip it to True.

Output is a single VAE, which you wire into VAEDecode just like any other VAE connection. Nothing else changes downstream - this is purely a loading-side swap.

Installing it

Same as the rest of the pack: ComfyUI Manager → search "ComfyUI-QuantOps", or:

cd ComfyUI/custom_nodes
git clone https://github.com/silveroxides/ComfyUI-QuantOps

Restart ComfyUI and let it install unifiedefficientloader>=0.5.2 from requirements.txt - that package is what makes the direct loading work, so if the node throws an import error, it's the first thing to check.

The caveat

The whole QuantOps pack is deprecated. The README is blunt about it: int8 ConvRot support is built into ComfyUI now, the author is done maintaining, and low-effort issue reports just get closed. So treat this node as a tool for a specific problem - a VAE that misbehaves under dynamic VRAM loading - rather than a long-term dependency. If your stock VAE loads fine, you don't need it. If it doesn't, this is one of the few genuinely useful survivors of the pack, because it sidesteps a ComfyUI loading path that still trips people up.