Eric Qwen-Image Diagnose LoRA

Eric Qwen-Image Diagnose LoRA answers a very specific question: is this LoRA even compatible with Qwen-Image? You give it a LoRA file, it inspects the file without applying it, and it returns a compatibility report as a string. If you've ever loaded a LoRA and gotten static instead of a subject, this is the node that tells you whether the file was the problem.

What the report actually contains

From the source, the diagnosis loads the .safetensors and checks its keys against the loaded Qwen-Image transformer:

• Adapter type - LoRA, LoKR, LoHa, or unknown.
• Rank and alpha - the standard LoRA hyperparameters, read from the file.
• How many adapter modules match the transformer architecture - the count of keys that line up with real layers.
• A verdict - COMPATIBLE, PARTIAL, or INCOMPATIBLE.

That verdict is the whole point. Qwen-Image has its own LoRA format and its own library - a LoRA trained for the Edit model won't apply to the generation pipeline, and an SDXL-era LoRA was never going to work on a 20B MMDiT no matter how good it looked on CivitAI. When something's wrong, this node distinguishes "wrong file for this model" from "file's fine, something else in your graph is broken."

The inputs

• pipeline - a QWEN_IMAGE_PIPELINE, from any generation loader. It's required because the check compares the LoRA's keys against the loaded transformer, so the pipeline has to be alive. (Which is also a hint: the diagnosis is architecture-level, not "is this a good LoRA" - it can't tell you if a compatible LoRA is badly trained, only whether it's the right shape.)
• lora_name - dropdown from ComfyUI/models/loras/, same list as Apply LoRA.
• lora_path_override - full path if your file lives outside the loras folder.

The output is a single report string. The node is marked as an output node and its IS_CHANGED returns NaN, meaning it runs every queue even if you don't do anything - that's deliberate, so the report is always fresh. Wire the report into a text display node to read it comfortably, or just watch the ComfyUI console, since the report also prints there.

When you'd actually reach for it

• A LoRA loads without error but produces noise → run this before you start messing with weights. If it says INCOMPATIBLE, you've saved yourself an hour of fiddling.
• You downloaded a "Qwen-Image LoRA" from a site with sloppy tagging → run this to confirm it's really for the generation model and not the Edit line.
• You're merging or training and want to sanity-check your output file → the rank/alpha readout is a nice free check.

It's the specialized sibling of Apply LoRA - the author clearly built it after fielding "why does my Qwen LoRA do nothing" questions, and it's the rare diagnostic node that's genuinely worth keeping in your test graph.

One honest caveat

The tooltip says it plainly: the LoRA is inspected but NOT applied. If the report says COMPATIBLE but your image still looks wrong, the problem is downstream - weight, staging (remember Apply LoRA has per-stage weights now, so a style LoRA at stage-1 weight 0 does nothing at all), or the LoRA simply being weak. Diagnosis can't save you from a bad weight choice, only from chasing a ghost.

Install: ComfyUI Manager search "Eric Qwen Edit", or git clone https://github.com/EricRollei/Eric_Qwen_Edit_Experiments into custom_nodes/ and restart. Needs the Qwen-Image pipeline loaded upstream - it's a reader, not a loader.