Eric Qwen-Image Apply LoRA

Eric Qwen-Image Apply LoRA loads a LoRA onto the Qwen-Image generation pipeline. Same idea as every LoRA node you've used, with one twist that makes it worth its own article: instead of a single weight, it exposes three per-stage weights (weight_stage1, weight_stage2, weight_stage3), because the generation nodes in this pack run progressive multi-stage pipelines. That's a genuinely different way to use a LoRA.

How the three weights work

The generation side of this pack (Generate aside) is staged: a low-res draft stage, a refinement stage, and a polish stage. This node lets your LoRA participate differently in each. The tooltips spell it out:

• weight_stage1 - draft/composition weight. Also the weight used by the non-UltraGen nodes. 0 = disabled.
• weight_stage2 - refinement weight. 0 = disabled.
• weight_stage3 - final polish weight. 0 = disabled.

A concrete pattern: a style LoRA you want to dominate the draft so composition locks in your direction, but fade on the polish pass so the detail stage doesn't over-apply it. Or the reverse - a realism LoRA you only want in the later stages where fine texture gets made. With a single-weight LoRA node you'd have to accept one strength for the whole run; here you get three.

The lora_name dropdown lists .safetensors files from ComfyUI/models/loras/, same as core ComfyUI. lora_path_override is the escape hatch if your LoRA lives somewhere else - leave it empty to use the dropdown.

Chaining and keeping

Multiple Apply LoRA nodes can be chained - the output of one feeds the input of the next - which is how you stack several LoRAs with independent weights, exactly like the README promises. And then there's keep_loaded (default True): keep the adapter loaded between runs so subsequent runs reuse it instead of reloading (faster for final renders); set it False if you're swapping LoRAs constantly and don't want stale adapters sitting around. The counterpart cleanup node is Eric Qwen-Image Unload LoRA, which strips everything and restores base weights.

The output is the same QWEN_IMAGE_PIPELINE type, so it drops straight into Generate, Multi-Stage, or UltraGen.

Realistic notes

Two things worth knowing. First, the weight range here is -2.0 to 2.0, not the 0-1 you might expect - the README says so, and negative weights are a real (if finicky) way to suppress a trained concept. Second, Qwen-Image responds unusually well to LoRAs per community consensus (the KB's own +852 thread calls it "very underrated" precisely because it "responds insanely well to LoRAs"), but you're still applying to a 20B model: keep keep_loaded True when iterating, because re-fusing adapters onto a 38 GB transformer is not instant.

The classic beginner trap is wiring a Qwen-Edit LoRA (the pose/angle/style library trained on the editor) into this node. Those LoRAs do nothing on the generation pipeline - the Edit line has its own library that doesn't transfer. If a LoRA loads fine but produces zero effect, check which model it was trained for first.

Install: ComfyUI Manager search "Eric Qwen Edit", or git clone https://github.com/EricRollei/Eric_Qwen_Edit_Experiments into custom_nodes/ and restart, with the Qwen-Image pipeline loaded by Eric Qwen-Image Loader upstream.