Unet Loader (LoraManager)

Modern models like Flux ship as split components - the diffusion model (the UNet) in one file, the text encoders and VAE separately. This node loads just the diffusion-model part, the same job as ComfyUI's core "Load Diffusion Model" node, but wired into LoRA Manager so it shows up alongside the rest of the pack's model handling and picks its file from the folders the manager tracks.

The reason it's in a LoRA pack at all: if you're running a Flux-family workflow with LoRAs from this manager, having the base UNet loader in the same family keeps everything consistent and lets the manager's browsing and metadata cover your checkpoints and diffusion models too, not just LoRAs.

How it works

You pick a diffusion-model file and a weight precision, and it hands back a MODEL ready for your sampler (after you patch LoRAs onto it). Note it only outputs the model - no CLIP, no VAE - because those are separate files in this kind of split setup; load them with their own loaders.

The weight_dtype choice is the one that matters for VRAM. Running the model in fp8 (8-bit float) roughly halves the memory the weights take versus the default, which is often the difference between a big model fitting on your card or not. The catch is a small quality/precision trade - fp8 is an approximation of the full weights.

Inputs and outputs that matter

• unet_name (enum, required) - which diffusion-model file to load, from your diffusion_models / unet folder. (The list is empty until you actually have models there.)
• weight_dtype (enum, required) - default (full precision, most VRAM, best fidelity), fp8_e4m3fn and fp8_e5m2 (two 8-bit formats that cut VRAM), or fp8_e4m3fn_fast (the faster fp8 path). Start with default; drop to an fp8 option if you're tight on memory.

One output: MODEL - feed it into your LoRA loader or straight to the sampler.

How to install it

Via ComfyUI Manager: search lora-manager, Install, restart. Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/willmiao/ComfyUI-Lora-Manager.git
cd ComfyUI-Lora-Manager
pip install -r requirements.txt

then restart. The pack itself is a light install; the heavy part is the diffusion-model files you point it at, which you download separately.

Common issues & troubleshooting

The dropdown is empty. No files in the diffusion-models folder yet. Put your .safetensors UNet/diffusion model where ComfyUI expects it (typically models/diffusion_models or models/unet) and refresh.

Out-of-memory on load. Switch weight_dtype to fp8_e4m3fn. If the model still won't fit, the diffusion model isn't the only thing on your card - text encoders and the VAE take memory too, and a GGUF-quantized build (loaded with a different node) goes smaller still.

Where's the CLIP and VAE? Not here - this loads only the model. That's correct for split-file setups; load the text encoder(s) and VAE with their own nodes.