UNETLoader_motorway_edition

Not every diffusion model ships as a full checkpoint with CLIP and VAE attached. Standalone UNet files - the diffusion_models folder is full of them - are just the denoiser, and they're what most modern video and image workflows load. UNETLoader is the stock node for that, and this _motorway_edition clone does the same load through the agilly1989 Motorway pipe: pick the file, optionally downcast it to fp8, and store the resulting MODEL in the shared dict by key.

The Motorway in one breath: start with MotorwayStart, then every _motorway_edition node swaps data wires for key strings, reading and writing a shared dict that travels down a single MOTORWAY connection. These clones are generated from ComfyUI's own node registry, so underneath this is the plain UNETLoader - same files, same formats, same behavior.

What it does. unet_name is the dropdown listing everything in models/diffusion_models. weight_dtype is where the interesting choice lives: default keeps the file's native precision, fp8_e4m3fn and fp8_e4m3fn_fast load it as fp8 (the _fast variant also enables fp8 optimizations for the attention path), and fp8_e5m2 is the slightly less accurate fp8 format. fp8 is basically "half the size, a hair less precise" - if the model fits in VRAM natively, just use default; if you're squeezing a 14B Wan onto a smaller card, fp8 is the lever people actually pull.

The inputs that matter. MOTORWAY is optional here - a loader has no data inputs, so connect an existing Motorway to drop the model into it, or leave it unconnected to act as a start. OUTPUT_MODEL_key defaults to MODEL (uppercase, which is what every consumer's INPUT_model_key field will expect if you retype it to match). The only output is the MOTORWAY port; the loaded model lives in the dict, not on a wire.

Common issues. The key case trap is the classic: this stores under MODEL, consumers default to model, so line those up or you'll get 'MotorwayClass' object has no attribute 'hash_' - the pack's "key not in the Motorway" error. Also, fp8 files can be slightly slower to load and marginally lower quality than native; the trade is VRAM, not speed. And the standing caveat: the pack's current main branch has the clone registration commented out in __init__.py, so if this node is missing from your menu, that's the version you have.

Install via ComfyUI Manager (search "ComfyUI_agilly1989_motorway") or git clone https://github.com/agilly1989/ComfyUI_agilly1989_motorway into custom_nodes, then restart. No models to download, no dependencies. Solo-dev beta pack - the Motorway is a tidy-up layer, and for a loader as straightforward as this one, it's a fine place to keep it.