Eric Qwen-Edit Apply LoRA

Qwen-Edit's stock behavior is fine, but the community's fine-tunes make it sing - the Lightning LoRA is the one that drops inference from 50 steps to 8, and style LoRAs (like PixelSmile for expressions) are what make the conditioning system really respond. Eric Qwen-Edit Apply LoRA loads those adapters into the pack's edit pipeline.

The catch with this pack is that it runs the real Hugging Face diffusers pipeline rather than ComfyUI's generic UNET abstraction, so the standard ComfyUI "Load LoRA" node can't touch it. This node is the diffusers-native path: it loads a .safetensors LoRA onto the pipeline's transformer and hands the modified pipeline on.

The inputs

• pipeline - from any pack loader.
• lora_name - dropdown of .safetensors files in ComfyUI/models/loras/. The same folder you already drop SD/Flux LoRAs into.
• weight - −2.0 to 2.0, step 0.05, default 1.0. 1.0 = full, 0.5 = half. Negative weights flip the adapter's influence.
• lora_path_override - optional full path to a LoRA outside the standard folder; leave empty to use the dropdown.
• keep_loaded - default true: load the adapter once and reuse it on subsequent runs (faster, good for final-render settings). Set false to unload all adapters and reload this LoRA every run - a clean state each generation, which is what you want when testing or swapping LoRAs frequently.

Output is the modified pipeline, ready to chain into an edit node or another Apply LoRA (multiple LoRAs chain with independent weights).

Practical notes

• Chain multiple Apply LoRA nodes to stack adapters with different weights.
• The README's tip: use the Lightning LoRA with 8 steps for fast iteration on every edit node; 50 steps without it. That's the single biggest speedup available in this pack.
• Qwen LoRAs are architecture-locked - the same rule as every modern base (a Flux LoRA won't load on Qwen-Edit). The Diagnose LoRA node in this pack will tell you whether a given file is even compatible before you commit to a render.

Installing it

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

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

Restart, download Qwen/Qwen-Image-Edit-2511 (~54 GB), load via the pack's loader. The pack pins diffusers>=0.34,<0.38 and transformers>=4.50,<5 - the transformers cap exists because 5.x removed HybridCache which older diffusers imports, and a broken import at startup is usually that mismatch.

Troubleshooting

• "No LoRAs in dropdown" - you haven't put any .safetensors in ComfyUI/models/loras/, or it's scanning the wrong folder. Use lora_path_override to point at a full path as a workaround.
• Output turned to noise / nothing happened - the LoRA is probably trained for a different architecture. Run Eric Qwen-Edit Diagnose LoRA on it; it reports adapter type, rank/alpha, module matches, and a compatibility verdict without applying anything.
• LoRA fighting with a direction edit - directions make the model more responsive; if an Apply-LoRA + direction combo overshoots, drop the LoRA weight to 0.5 or lower the direction scale.
• Swapping LoRAs between runs leaves residue - that's what keep_loaded = false is for; it forces a clean reload every run.