Eric Qwen-Edit Diagnose LoRA

Somewhere on the internet there is a LoRA that claims to work with Qwen-Edit and doesn't. It loads, it runs, and it produces static. Eric Qwen-Edit Diagnose LoRA is the tool for that exact moment: it opens a LoRA file, inspects its keys without applying it, and tells you what it actually is and whether the Qwen-Edit transformer can use it.

The report covers the four things you actually need to know:

• Adapter type - standard LoRA, LoKR, LoHa, or unknown.
• Rank and alpha - the values the file was trained with, which tell you how much capacity the adapter has.
• How many adapter modules match the transformer architecture - the key number. A LoRA's key names must line up with modules in the model; mismatches mean it was trained for a different architecture.
• A verdict - COMPATIBLE, PARTIAL, or INCOMPATIBLE.

The inputs

• pipeline - from any pack loader (needed so it can compare the LoRA's keys against the actual loaded transformer).
• lora_name - dropdown of LoRAs in ComfyUI/models/loras/.
• lora_path_override - optional full path for a LoRA outside the standard folder.

Output is report (STRING) - printed to the console and returned as a text value you can display in the UI.

Why this is worth doing

Modern architecture lock-in is brutal: every base model needs its own LoRAs, and nothing checks whether a file was made for Qwen-Image-Edit vs Qwen-Image vs some unrelated DiT until you run it. A noisy output after a 50-step render is a slow way to find out. Diagnose answers "is this file even for this model?" in a second, no GPU burn. It's also the right first step when a LoRA fails to load with a confusing diffusers error - the key-matching count tells you whether it's a structural mismatch or something subtler.

Installing it

ComfyUI Manager → "Eric Qwen-Edit", or:

cd ComfyUI/custom_nodes
git clone https://github.com/EricRollei/Eric_Qwen_Edit_Experiments.git

Restart, load Qwen/Qwen-Image-Edit-2511 (~54 GB) via the pack's loader. The pack pins diffusers>=0.34,<0.38 and transformers>=4.50,<5.

Gotchas

• "PARTIAL" isn't "no" - a partial match can still work if the matched modules carry the meaningful weights; it can also mean a half-applied adapter that makes things weird. Render one low-step test (Lightning LoRA speed, 8 steps) before trusting it.
• This checks structure, not quality. A COMPATIBLE verdict means the keys line up, not that the adapter is good. A badly-trained LoRA of the right architecture still diagnoses as compatible - diagnose finds the wrong-file problem, not the bad-training problem.
• It does not apply the LoRA. Nothing is loaded or modified; the pipeline comes back untouched. That's deliberate - this is the inspection node, Apply LoRA is the application node.
• Diagnosing needs the pipeline loaded, because the module-match count compares against the real transformer. You can't diagnose with an empty graph.