UNet loader with Name (Image Saver)

This is ComfyUI's "Load Diffusion Model" node with a bonus output: the model's filename as a string. It exists for the same reason the pack's checkpoint loader does - the Image Saver nodes need you to tell them exactly which model ran, and the least error-prone way to get that name is directly from the loader. Use this when your model comes as a standalone diffusion/UNet file rather than a full checkpoint, which is most of the Flux-era and quantized world now.

Load the model, get MODEL for your sampler, and get its name for your metadata in one node. That's the whole pitch, and it's a good one if you post to Civitai.

How it works

It loads a diffusion model from your models/diffusion_models (a.k.a. unet) folder and outputs the MODEL, plus the filename you picked. The one real setting is weight_dtype, which controls how the weights are cast in memory. Leaving it on default uses the file's native precision; the fp8_e4m3fn / fp8_e5m2 options load the model in 8-bit float to save VRAM. fp8_e4m3fn_fast trades a hair of quality for speed on GPUs that support it. This is the same fp8 lever people reach for on big Flux-class models - worth knowing that fp8 is a different quantization path from GGUF's Q-ladder; if your model is a .gguf file this isn't the loader for it.

The inputs and outputs that matter

• unet_name (required) - dropdown of diffusion model files you have installed.
• weight_dtype - default, fp8_e4m3fn, fp8_e4m3fn_fast, fp8_e5m2. Start with default; drop to an fp8 mode only if you're tight on VRAM.

Outputs:

• model - into your sampler / model-patching chain.
• filename - into the modelname of an Image Saver or Image Saver Metadata node, so the saved metadata names the real file.

Installing it

ComfyUI Manager: search ComfyUI Image Saver, install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/alexopus/ComfyUI-Image-Saver
cd ComfyUI-Image-Saver
pip install -r requirements.txt

Restart, hard-refresh. No model downloads from the pack itself - it loads what's already in your diffusion_models folder.

Where people get tripped up

The most common confusion is loader vs. file type. This node is for single-file diffusion models (the Flux-style .safetensors UNet files). If your model is a full checkpoint with CLIP and VAE baked in, use Checkpoint Loader with Name. If it's a GGUF-quantized file, you need a GGUF loader from a different pack - this one won't list .gguf files, and that's expected, not a bug.

Second: fp8 isn't free quality. On models that were trained or released in fp8 it's basically lossless, but forcing an fp16 model down to fp8 can cost a little fidelity - if you have the VRAM, default is the safe choice. And as with the checkpoint loader, actually wire the filename output into your saver; otherwise you've gained nothing over the stock diffusion loader.