⬡ Polyhedron Load Model

The ⬡ Polyhedron Load Model replaces the loader plus a separate model-picker switch with a single node that is both. You fill up to six slots with model files, use the select widget to say which one is live, and the node loads that pick itself - there's no separate loader node downstream, and no orphaned selector whose wiring you have to keep in sync.

The useful part for people juggling checkpoints: the file dropdowns don't just list one folder. Each slot spans unet, diffusion_models, checkpoints, and both GGUF folders in one list, so fp8 and GGUF quantized models sit next to full checkpoints and you compare them at a glance. The README sums up the sizing philosophy: a second slot appears only once the first is filled, so the node stays exactly as small as the job.

How it works

The slots are alternatives, not a stack - exactly one model is loaded per run, which is what a selector should do. The select widget's upper bound follows how many slots are filled, so it can't point at an empty one; as soon as you fill the last slot in use, a fresh one appears. When you flip select, the node loads the new file and the old one is dropped, which keeps VRAM from quietly accumulating every model you've ever selected.

weight_dtype is the same safetensors weight-dtype override you get on core's UNETLoader (default, fp8_e4m3fn, fp8_e4m3fn_fast, fp8_e5m2) - the "how much precision am I trading for VRAM" knob. It's ignored for GGUF files, which are already quantized; you pick the Q4/Q5/Q8 you want at the file level instead.

What comes out

• model (MODEL) - the real loaded model. This is the important one: no filename-string shenanigans, it's a genuine MODEL that wires into samplers and the rest of the suite.
• model_name (STRING) - the selected filename, for Save-node prefixes, notes, or foreign combos that want to know which model produced this render.
• info (STRING) - a readout: file | source kind | dtype | size. Wire it into a ⬡ Polyhedron Note for a live "what's loaded" panel on the canvas.

The model_name output is the quiet killer feature - it means your output filenames can say which model made the image, automatically.

Installing it

Part of the ⬡ Polyhedron Suite pack. ComfyUI Manager → "Polyhedron Suite", or:

cd ComfyUI/custom_nodes
git clone https://github.com/PolyhedronAI/ComfyUI-PolyhedronLoRAStack.git
# restart ComfyUI

No extra dependencies. GGUF support rides on whatever GGUF loader you already have in your setup.

Gotchas

The one mental shift: this node is the loader, so don't wire it into a downstream UNETLoader expecting a handoff - model goes straight into the LoRA stack or sampler. And remember the slots are exclusive; if you're building a dual HIGH/LOW Wan graph you'll use two of these nodes (or the pack's Select Model Switch idiom), not six slots on one. If a Polyhedron node shows up blank under ComfyUI's Nodes 2.0 renderer, disable Modern Node Design in Settings - the data is safe, only the drawing breaks.