LoRA Reporting

LoRA Reporting is the node that opens up a LoRA file and tells you what's inside - training tags, key count, parameter count, and the trainer's metadata. It does the same "doctor the file" trick as the pack's LoRA Loader (Tags), but as a pure inspection tool: no model gets modified, nothing gets loaded into the graph, you just get a report.

If you've ever downloaded a LoRA and wondered what it was trained on, at what resolution, with how many steps, or whether the top training tags match what the author promised - that's this node. Most of that info is written into the file by kohya-style trainers as ss_* metadata keys, and this node is the least-effort way to read them in ComfyUI.

The inputs

• lora_name - dropdown of your ComfyUI/models/loras folder. Pick a file; that's it.
• report - the mode:
  • tags (default) - the top training tags by frequency, from ss_tag_frequency. Limited by limit. This is the same data the LoRA Loader (Tags) returns as a string, but for a LoRA you aren't necessarily loading.
  • size - key count, total parameters, and the interesting ss_* training metadata: resolution, epochs, steps, network args, noise offset, and so on. This is the "spec sheet" mode.
  • details - the full metadata dict dumped as JSON, including everything a trainer wrote that size doesn't filter to. Raw and verbose; use it when you need a specific obscure key.
• limit (default 10) - how many tags in tags mode.

The outputs

• STRING - the report text. Wire it to a text display node to keep it on the canvas, or read it in the node output.
• IMAGE - a declared second output that the current implementation doesn't fill for LoRA reports; it's a placeholder slot, so don't build around it.

When you'd actually use it

Two good habits. Before you load an unknown LoRA, run size to see its resolution and training setup - that tells you whether it's aimed at the checkpoint you're about to stack it on. And when a LoRA needs trigger words you can't find on its page, run tags and read the top few off the file. It's also a neat sanity check on whether a LoRA was tag-trained at all: if tags comes back empty, the file has no ss_tag_frequency, and the LoRA Loader (Tags) node will hand you an empty string too.

Install and gotchas

ComfyUI Manager (search "DareMerge") or:

cd ComfyUI/custom_nodes
git clone https://github.com/54rt1n/ComfyUI-DareMerge

Restart ComfyUI. Dependencies are matplotlib, numpy, torch, pillow - no model downloads (and this node barely touches the heavy ones; it's mostly reading a header).

The main thing to keep expectations honest about: LoRA files only contain what their trainer wrote. A LoRA trained with a minimal config may have almost no metadata, in which case size and details are thin and that's not a bug. And ss_tag_frequency reflects training data, not curated trigger words - useful, but read the top tags as "what this model saw a lot," not "the guaranteed activation phrase."