Load UNet Model with Name JNK

The name is a lie, and you should know it before it costs you an afternoon. Load UNet Model with Name does not load a UNet. Open the source and it's the same one-liner as the LoRA version - os.path.splitext(os.path.basename(model))[0] - a filename strip and nothing more. The "model" output is a string (the full filename), not a loaded MODEL, and the "name" output is that same filename with folder and extension removed. No weights, no VRAM, no diffusers, no nothing.

It's a bookkeeping node wearing a loader's costume, and the confusion is the whole problem. The README frames it as "Return UNet model and model name," which doesn't help - but the code doesn't lie: both outputs are text. What it is genuinely good for is the same job as its siblings: knowing which UNet produced an image. Wire the first output into a real Load Diffusion Model node (the core one that actually loads a diffusion model), feed name to your save nodes, and your filenames carry the UNet identity automatically.

Inputs and outputs

• model - dropdown of everything in ComfyUI/models/diffusion_models/, subfolders included.

Outputs:

• model - the filename string (e.g. flux1-dev-fp8.safetensors). Feed this into the core Load Diffusion Model node's model slot.
• name - basename without extension (flux1-dev-fp8). The one you use in filenames.

Installing it

Pack install: search JNK in ComfyUI Manager, or

cd ComfyUI/custom_nodes/
git clone https://github.com/Aljnk/ComfyUI-JNK-Tiny-Nodes.git

No dependencies - it's a string operation.

Where people get burned

The failure mode is a type mismatch: try to wire the first output straight into a KSampler or a model-mixing node and ComfyUI rejects the connection, because it's a string and not a MODEL. That's your clue you've been fooled. The correct mental model: this node is a name picker that doubles as a dropdown, and you still need the core Load Diffusion Model node to actually pull the UNet into memory. If that sounds redundant - it is, deliberately. The "with Name" family trades one extra node for self-documenting output files. Compare it with Load Checkpoint Model with Name, which is the one member of the family that genuinely loads its model, and you'll see the difference instantly. When you just need the name, this is the cheap option; when you need an actual loaded UNet, look elsewhere.