Model Info Inspector 🔬

Download a checkpoint from somewhere, drag it into ComfyUI, and the first question is always the same: what the hell is this, and will my VRAM survive it? Model Info Inspector 🔬 answers that before you commit to a full load. It reads a checkpoint's raw state dict and reports the key count, an architecture guess, the weight precision, and an estimated parameter count - without constructing a MODEL/CLIP/VAE.

The "without loading it" part is the whole point. ComfyUI's normal loader builds the full model graph, which costs real time and VRAM, and it fails on checkpoints it can't match to a config. This node deliberately skips that: it reads the safetensors/state-dict structure only, so it works even on files ComfyUI can't fully load, and it's fast enough to run on anything in your checkpoints folder.

How it works

It pulls the list of files from ComfyUI's checkpoints/ search paths (so it only sees checkpoints ComfyUI knows about - same registry as the normal loader), opens the chosen one's raw state dict, and inspects key names. The architecture guess is signature-based: double_blocks keys suggest FLUX, joint_blocks suggest SD3, and so on down the family tree. Precision and parameter count come straight from the tensor dtypes and shapes.

The inputs and outputs

One input: ckpt_name, a dropdown of your available checkpoints. Outputs are four: summary (a readable report), key_count (INT), architecture (STRING - the guess), and precision (STRING - the detected weight dtype). Those last three are individually wireable, which matters if you want to log or compare them programmatically.

Where you'll actually use it: auditing a checkpoint you don't trust or remember, checking whether a file is fp16 vs fp8 before deciding if it fits your card, and sorting a folder of mixed merges without loading each one. It pairs naturally with the pack's VRAM / Model Size Estimator for a full "is this loadable" check before you queue.

Installing it

Part of OmniNodes:

cd ComfyUI/custom_nodes/
git clone https://github.com/TensorVizion/OmniNodes

Restart ComfyUI (ComfyUI Manager: search "OmniNodes"). No extra dependencies - it uses folder_paths, comfy.utils, and torch, all bundled. Registers under TensorVizion/Model Utilities. If it doesn't appear, restart completely and check for [OmniNodes] ✅ Loaded in the terminal.

One honest note: the architecture guess is a heuristic based on key-name signatures, not a verified config match. For the common families it's reliable; for exotic or heavily-modified checkpoints, treat it as a strong hint rather than gospel - which is fine, because the point is to know before you load, not to write a spec sheet.