Nunchaku Qwen Image LoRA Stack V1

If you've used the Power Lora Loader from rgthree, you already know the interface here. V1 is this pack's take on that clean, minimalist row layout, one line per LoRA with a toggle, a name and a strength, brought over to Nunchaku-quantized Qwen-Image models. It came in through a community PR (from avan06, inspired directly by rgthree's Power Lora Loader), and the maintainer's own note is that they'd wanted a UI this tidy for ages but couldn't build it themselves. So: the ergonomics of the LoRA stacker everyone loves, wired to work on 4-bit Nunchaku models where the original can't.

The why, briefly: Nunchaku runs Alibaba's 20B Qwen-Image in 4-bit SVDQuant so it fits on consumer VRAM, and stock LoRA loaders can't patch those quantized weights. This pack is the bridge, and V1 is the nicest-looking way to stack several LoRAs across it.

How it works

The LoRA rows live on the node face as interactive widgets, the same way rgthree's loader works, so you add rows, toggle each on or off, and dial strengths right there without rewiring anything. Under the hood it maps every enabled LoRA onto the quantized model and passes the patched model to your sampler.

The controls the node exposes as actual inputs are the model-level ones:

• model (MODEL) - your Nunchaku Qwen-Image model.
• cpu_offload (auto / enable / disable, default disable) - offload to RAM to save VRAM if you need to.
• apply_awq_mod (default true) - this one's worth knowing. Nunchaku AWQ-quantizes the model's modulation layers (img_mod.1 / txt_mod.1), and naively pushing LoRA onto them used to produce noise. The pack applies a runtime patch to fix that, and V1 turns it on by default. Leave it on unless you're chasing a specific artifact.
• stack_enabled (default true) - a master switch for the whole stack, handy for an A/B against the base model.

Output is one patched MODEL.

A real gotcha with this one

V1 uses the rgthree-style JavaScript UI, and that UI does not behave properly under ComfyUI's Nodes 2.0 canvas. Use the standard (LiteGraph) canvas with V1. If you're stuck on Nodes 2.0, pressing F5 to refresh will make your row changes register, but honestly the cleaner move is to switch to the legacy canvas for this node, or use V3 instead, which was built for Nodes 2.0. This is the same compatibility story rgthree's own nodes have been living through since the late-2025 frontend rewrite, so if you know that saga, this is a familiar shape.

Installing it

Via ComfyUI Manager (search the pack, install, restart) or by cloning:

cd ComfyUI/custom_nodes
git clone https://github.com/ussoewwin/ComfyUI-QwenImageLoraLoader

then restart. It rides on the official ComfyUI-nunchaku plugin, which you install and get working first, wheel matched to your PyTorch/CUDA. Python 3.11+.

Where people get burned

• Nodes 2.0. Covered above, but it's the number-one thing that makes this node look broken. Wrong canvas, mangled rows. Switch to LiteGraph.
• LoKr LoRAs are skipped. Only Standard and PEFT formats apply on Nunchaku models. If a toggled-on row does nothing, check the format.
• Don't stack a lightning LoRA on an already-distilled model. It's a reliable route to black images. One distillation on top of another fights itself.