FL Model Inspector

You've got a .safetensors or .ckpt file with a filename that tells you nothing - no architecture, no precision, no idea what's actually inside until you load it and something either works or explodes. FL_ModelInspector is the node for figuring that out ahead of time: point it at a checkpoint path and it reads the file's structure and hands you back a readable report, without you needing to drop into a Python shell and poke at state_dict keys yourself.

How it works

Unlike most nodes in this pack, it doesn't take an already-loaded MODEL as input - it takes a path string and loads the checkpoint itself, with fallback loading methods so a slightly unusual file format is less likely to just fail outright. From there it extracts parameter count, layer shapes, and does some architecture analysis to characterize the model's structure. You control how deep the report goes: include_layer_details breaks down individual layer shapes (on by default), include_state_dict dumps the raw state dict info (off by default - it's a lot of output), and show_all_keys lists every single key in the checkpoint rather than a summarized view.

The inputs and outputs that matter

• ckpt_path - the file path to inspect. This is the only real input; everything else just controls verbosity.
• include_layer_details, include_state_dict, show_all_keys - three booleans trading detail for report length. Start with just include_layer_details on and only flip the other two if the summary doesn't answer your question.
• Outputs: model_info (STRING) is the actual report - read it, or wire it into FL_ShowText/a save-to-file node if you want it somewhere persistent. model (MODEL) is the loaded checkpoint passed through, meaning this node can double as your checkpoint loader - you get the inspection report and a usable model out of the same node, instead of loading it twice.

How to install it

Via ComfyUI Manager: search "ComfyUI_Fill-Nodes" and install. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/filliptm/ComfyUI_Fill-Nodes

then restart ComfyUI. No extra dependency beyond PyTorch's own checkpoint-loading tools, which ComfyUI already needs anyway. The pack overall is a large, general-purpose collection from one prolific solo author, filliptm - spanning image FX, PDF tools, GPT/Gemini/Fal API wrappers, and an entire KSampler family - but this specific node needs nothing beyond what a normal ComfyUI install already has.

Common issues & troubleshooting

Report is huge and hard to read. That's usually include_state_dict or show_all_keys turned on for a large model - both dump a lot of raw detail. Turn them off and start from the summarized include_layer_details view; only reach for the full dump when you're hunting a specific key.

Loading fails on an unusual checkpoint. The node has fallback loading methods precisely because checkpoint formats aren't perfectly standardized across the ecosystem - but a genuinely corrupted or partially-downloaded file will still fail. If a normal checkpoint loader also can't open the file, that's your answer: it's not this node, it's the file.

You just want to know if a checkpoint is fp16 vs fp32, or roughly how big it is. The parameter count and layer-shape summary is exactly what to look at - you don't need show_all_keys for that, and turning it on will just bury the answer in noise.