Hello, I am xiaozhijason aka lrzjason. I created a helper nodes which could load any models in any place of your workflow.

Github: https://github.com/lrzjason/ComfyUI-LoaderUtils

✨ Enter ComfyUI-LoaderUtils: Load Models Only When Needed

I created a set of drop-in replacement loader nodes that give you flexible model loading. How? By adding a magical optional any parameter to every loader – letting you sequence model loading based on your workflow’s actual needs.

Key innovation:
✅ Strategic Loading Order – Trigger heavy models (UNET/Diffusion model) after text encoding
✅ Zero Workflow Changes – Works with existing setups (just swap standard loaders for _Any versions and connect the loader before it need)
✅ All Loaders Covered: Checkpoints, LoRAs, ControlNets, VAEs, CLIP, GLIGEN – [full list below]

💡 Real Workflow Example

LoaderUtils

1. Run text prompts & conditioning

2. Then load UNET via UNETLoader_Any

3. Finally load VAE via VAELoader_Any after sampling → Stable execution on 8GB GPUs ✅

🧩 Available Loader Nodes (All _Any Suffix)

Standard Loader

Smart Replacement

CheckpointLoader

→ CheckpointLoader_Any

VAELoader

→ VAELoader_Any

LoraLoader

→ LoraLoader_Any

ControlNetLoader

→ ControlNetLoader_Any

CLIPLoader

→ CLIPLoader_Any

(+7 more including Diffusers, unCLIP, GLIGEN, etc.)

No trade-offs: All original parameters preserved – just add connections to the any input to control loading sequence!