Gen2 Load QwenImage LoRA (outdated)

LoRA handling inside this pack's (outdated) QwenImage pipeline works a little differently from ComfyUI's usual LoRA loader, and it's worth knowing before it bites you. Rather than patching the model at load time, this node is a collector: it records which LoRA files you want and at what strength, then the Gen2 QwenImage Control Sampler actually merges them when it runs. That's a deferred-loading design, and it's exactly what makes the node chainable - you can daisy-chain several of them and the sampler applies them in order, as a stack.

Inputs and output

• lora_name - an enum of LoRAs from your ComfyUI loras folder.
• strength - the merge strength, default 1.0, range −100 to 100 (step 0.01). Negative strengths are allowed, which is the usual "subtract this LoRA" trick.
• lora (optional) - a GEN2_LORA from a previous Gen2 Load QwenImage LoRA node. Wire this in to append to the stack instead of starting a new one.

Output: lora (GEN2_LORA) - a dict carrying lora_paths and lora_strengths lists. Feed it into the sampler's optional lora input.

So the stacking pattern is: Load LoRA A → Load LoRA B (with A's output wired into B's lora input) → sampler. The sampler then merges them in sequence at their recorded strengths. The order matters, and since the nodes preserve insertion order, you control it by how you chain.

Installation

Like the rest of the QwenImage section, this needs the full VideoX-Fun stack. From the README:

• VideoX-Fun as a custom node,
• diffusers (and transformers) in your environment,
• optionally ComfyUI-GGUF for quantized weights.

Install the pack via ComfyUI Manager (search "ComfyUI-gen2"), or:

cd ComfyUI/custom_nodes
git clone https://github.com/petmycat/ComfyUI-gen2
cd ComfyUI-gen2
pip install -r requirements.txt

Missing VideoX-Fun → [Gen2] QwenImage nodes not available at startup; the rest of the pack loads independently.

Common issues

• "LoRA not found" warning in the console but no error. The loader only appends to the stack if folder_paths.get_full_path("loras", lora_name) resolves. If the file moved or the enum is stale, it silently logs and continues with an empty stack - the sampler then runs without your LoRA. Watch the console.
• LoRA has no effect. The merge happens in the sampler, not here. If the sampler isn't wired with your lora output, this node did nothing. Also, QwenImage LoRAs are architecture-specific - an SDXL or Flux LoRA won't apply meaningfully to QwenImage.
• Stack order confusion. If you chain B into A but want A applied after B, you've wired the chain backwards. The stack is built in the order the nodes are chained.

One more honest note, same as the other "(outdated)" nodes: this is backward-compatibility plumbing. If you're running a modern native QwenImage workflow you almost certainly don't need this section at all. It exists to keep old VideoX-style pipelines reproducible, and it does that job fine.