Model Metadata Reader

ModelMetadataReader is the basic node of the three that ship in Sinphaltimus/comfyui_fedcoms_node_pack, and it's the one you'll actually reach for. Its job: read the metadata that lives in a model file and hand it to you as text. Why care? Checkpoints and LoRAs carry a surprising amount of useful junk in their headers - training tags, trigger words, the ss_* keys from a LoRA training run, the base model it was built on. That's all readable without loading a 7GB file into VRAM, and ComfyUI's default loaders won't show you any of it.

How it works

The node looks at the file extension and picks a method:

• .safetensors → opens with safetensors.safe_open() and returns the header metadata. This is the fast, genuinely useful path - only the header is read, the weights never touch RAM.
• .ckpt, .pth, .pt, .bin → loads the whole file with torch.load() and returns the top-level keys (usually just state_dict).
• anything else → "Unsupported model format".

The one input and one output

There's a single required input, model_path - a STRING with the placeholder "Enter full model path here". It's exactly what it sounds like: the full absolute path to the model file, typed or pasted in. There's no file browser. The output is one STRING: a timestamped log plus the extracted metadata as pretty-printed JSON. Wire it into any text display node, or feed it to a node that builds filenames or tags from what it finds.

Install

Three nodes, one tiny repo. Via ComfyUI Manager, search "comfyui_fedcoms_node_pack", or:

cd ComfyUI/custom_nodes
git clone https://github.com/Sinphaltimus/comfyui_fedcoms_node_pack

Restart ComfyUI. No model downloads, no heavy dependency. The requirements.txt is mostly the Python standard library plus torch, safetensors, and chardet - and torch and safetensors are already in every ComfyUI install, so in practice only chardet might be new. Manager installs it automatically.

Where people get burned

• It needs a real path. A relative path or a typo gets you ❌ Error: Model not found and nothing else. The pack ships a PowerShell script (PSscript/Model_Lister_with_paths.ps1) that walks your models folder and writes every full path to a model_list.txt for copy-pasting - Windows only, and clearly how the author actually runs this.
• It's not a universal reader. .onnx and .gguf files get "Unsupported model format." The "Enhanced" sibling handles ONNX, and the "Advanced" one claims the rest.
• The .ckpt path has a modern-PyTorch trap. Since PyTorch 2.6, torch.load() defaults to weights_only=True, which refuses to deserialize anything but plain tensors - a whole class of checkpoints now errors with "Weights only load failed." And it's worth remembering why safetensors exists: .ckpt is pickle, the format that can execute code when loaded. You're not in danger from merely reading keys here, but this node's ckpt support is best-effort, not something to rely on.

If what you care about is safetensors - LoRAs especially - this is the honest pick of the three. It's fast, and the metadata it returns is real.