LoRA Metadata Diff 🆚

Your LoRA folder has grown to 200 files and some of them are from trainers that wrote sloppy metadata. LoRA Metadata Diff 🆚 is the node that catches the embarrassing ones before they ship: it loads two LoRA files and diffs their structural metadata - rank, alpha, target modules, key and parameter counts - plus whatever embedded safetensors header metadata they carry (base model tag, trigger words, training params).

The name undersells it slightly. It's not comparing the learned weights themselves; it's comparing the bookkeeping around them, which is exactly where LoRAs silently go wrong. "These two were trained on different base models" and "one of these lost its trigger-word metadata" are both diffs you want to catch before you upload a batch to CivitAI or hand a file to someone.

How it works

Both LoRA names come from ComfyUI's loras/ registry (the same dropdown source as the core LoRA loader - so files must be where ComfyUI looks, usually ComfyUI/models/loras/). The node reads each file's tensor structure to derive rank, alpha, and target-module prefixes, and reads the safetensors header for embedded metadata. The diff_report string lays out the two side by side and flags keys present in only one file, or present in both with different values. A structurally_compatible BOOLEAN gives you the bottom line - if it's False, these two live in different worlds and stacking them may not do what you expect.

The inputs and outputs

Two required inputs: lora_name_a and lora_name_b, both dropdowns of your installed LoRAs. Outputs are diff_report (STRING) and structurally_compatible (BOOLEAN). Wire the report into a text display; wire the boolean into a conditional if you want the workflow to branch on whether a pair is safe to combine.

It's the two-file version of the pack's LoRA Info Inspector, and it's categorized under TensorVizion/Model - the pack's known quirk that a handful of model nodes land in a separate submenu from Model Utilities. Not a bug, just where to look in the search list.

Installing it

Same pack, same routine:

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

Restart ComfyUI (or find "OmniNodes" in ComfyUI Manager). No extra dependencies beyond what ComfyUI ships. If it doesn't show up, restart completely and check the terminal for [OmniNodes] ✅ Loaded lines.

The honest limitation: it reports what's in the headers, and not every trainer writes useful headers. Two LoRAs that both say "SDXL base" can still behave differently in practice - but catching the mismatched-base-model case before it wastes an afternoon is more than worth the install.