Checkpoint Evidence Inspector

Ever stared at a 12GB checkpoint and wondered whether it's the RAW or Turbo variant of something, or whether that "official" file you grabbed actually matches the model card? Loading it to find out costs you VRAM and a few minutes. The Checkpoint Evidence Inspector reads only the bounded header of a local .safetensors file - no tensor payloads, no model load - and reports what's inside.

Safetensors files store their metadata (tensor names, shapes, dtype, sometimes even formatting metadata) in a header at the start of the file. That header is cheap to read and usually enough to fingerprint a model family. This node leans on that: it inspects the header, emits a report of what the checkpoint claims to be, and - because this pack cares about Krea 2 provenance - gives an advisory suggested variant. Read that word carefully: advisory, not confirmation. The README is blunt that the suggested variant is not authoritative. A file can have a misleading name or sparse metadata; the header tells you evidence, not truth.

What you get

The input is a single dropdown, checkpoint, listing the local model files ComfyUI can see (if none are present, it shows <no local safetensors available>). The output is one string, checkpoint_evidence, containing the header report as JSON.

It's an output node in the inspect-and-read sense: you wire it into nothing downstream, you just read the report. If you're doing forensic "is this the file I think it is" work before building a workflow, run it first.

Installing

Same pack install as every Sigmax node:

cd ComfyUI/custom_nodes
git clone https://github.com/rookiestar28/ComfyUI-Sigmax comfyui-sigmax

Restart ComfyUI, then find it under the Sigmax menu. ComfyUI Manager works too (search "ComfyUI-Sigmax"). Python 3.10+, ComfyUI 0.29.0+, no extra dependencies.

Where it fits, and where it doesn't

Where it fits: before you commit a workflow to a model, or when you have two checkpoints with confusing names and want a quick structural tell. Where it doesn't: it won't tell you a model is good, it won't validate quality, and it won't confirm a variant with certainty. It's the fast pre-flight check in a pack whose whole pitch is explicit, traceable model selection. If you want a deeper per-model look that actually inspects a loaded Krea 2 model, Sigmax's ProfileInspector is the sibling node for that. Use the evidence inspector as the cheap gate before spending the VRAM.