(Down)load UNET

The last of the pack's five download nodes, and the one aimed at newer-style checkpoints. Models like Flux ship their diffusion backbone as a separate file from CLIP and VAE, rather than one bundled .safetensors - ComfyUI loads that piece with its own "Load Diffusion Model" node, reading from models/diffusion_models (also referred to as UNET in older terminology). DownloadUNET fetches that specific file type from CivitAI or HuggingFace by ID.

How it works

Same pattern as the rest of this pack's download family: choose a source, provide a model_id, and it downloads the file into the right folder, returning the filename as unet_name. It's an output node, meaning it executes as part of the graph run even if you don't wire anything downstream of it - enough to use it just to pre-fetch the file.

The inputs and outputs that matter

• source - civitai or huggingface.
• model_id (STRING) - the CivitAI model ID or the HuggingFace repo path.
• base_model - SD1.5, SDXL, Flux.1, Kolors, or Pony. Given this node is really only relevant for architectures that split the diffusion backbone out separately, you'll mostly see this set to Flux.1 in practice.
• version_id (optional) and file_names (optional, multiline) - pin a version, or a specific file, when a model page offers a choice - worth knowing that UNET-only releases are often published in more than one precision (fp16, fp8, and various GGUF quant levels all showing up as separate files under one version), so file_names is where you'd steer toward the one that actually fits your VRAM.

Output: unet_name - wire into a diffusion-model loader node.

How to install it

Like the other Download* nodes, this comes from the bundled ComfyUI-Model-Downloader submodule, not the pack's main repo:

cd ComfyUI/custom_nodes
git clone --recursive https://github.com/liuqianhonga/ComfyUI-QHNodes.git

Or, on an existing non-recursive clone:

cd ComfyUI/custom_nodes/ComfyUI-QHNodes
git submodule init && git submodule update

If you installed via ComfyUI Manager and this node is missing, run the submodule commands above inside the resulting folder - Manager's plain-clone install doesn't reliably bring submodule-based node packs across in full.

Common issues & troubleshooting

Downloaded file is enormous or won't load in your VRAM budget. Full-precision UNET/diffusion-model files for current-generation checkpoints are large. If you didn't specify a file via file_names, you may have gotten the biggest precision variant on offer rather than an fp8 or GGUF-quantized one that would actually fit your GPU - check what files a version actually publishes before assuming you have no smaller option.

Node loads it but the rest of the workflow errors out. A standalone UNET file needs to be paired with the matching CLIP/text-encoder and VAE for that architecture - Flux's diffusion model doesn't work with an SDXL CLIP or VAE. Make sure the rest of your loading chain matches the base_model you picked here.

Gated repos return nothing. No API key or token field exists on this node, same as its siblings - gated HuggingFace repos and login-gated CivitAI downloads won't authenticate through it.