Nunchaku Qwen Image LoRA Loader

Here's the problem this node solves. You're running Qwen-Image through Nunchaku so the 20B model actually fits on your card, you've got a LoRA you want to add, and you drag in ComfyUI's stock LoraLoader... and it does nothing, or it errors. That's expected. Nunchaku stores the model as 4-bit quantized weights, and a normal loader has no idea how to patch those. This node does. It's the single-LoRA loader for Nunchaku Qwen-Image, and it's the one to start with.

Quick context on the two moving parts. Qwen-Image is Alibaba's 20B Apache-licensed image model, capable but genuinely heavy. Nunchaku is MIT HAN Lab's inference engine that quantizes diffusion transformers down to 4-bit (SVDQuant), which is how you run a 20B model fast on consumer VRAM instead of watching it swap to disk. The Nunchaku Qwen-Image release landed in August 2025 and the community jumped on it hard. The one thing it didn't ship cleanly was LoRA support, and that gap is exactly what this pack (extracted and modified from GavChap's nunchaku fork) fills.

How it works

A LoRA is a small adapter that patches a checkpoint's weights without retraining it, so a loader has to sit between the model and the sampler and intercept it on the way through. This node takes your loaded Nunchaku Qwen model, maps the LoRA's weight keys onto the quantized layers, applies them at the strength you set, and hands the patched model back out. Same model object, modified behavior, then on to your sampler.

The inputs and outputs that matter

Four inputs, one output, and only two you'll really touch:

• model (MODEL) - the Nunchaku Qwen-Image model, straight from your Nunchaku Qwen Image DiT Loader.
• lora_name - the dropdown of your installed LoRA files. Standard (Kohya/diffusers) and PEFT-format LoRAs work; LoKR, LoHa and IA3 do not (more on that below).
• lora_strength (default 1.0) - how hard the LoRA pushes. The old "keep it around 0.7" habit is SDXL-era advice; on modern bases people often run 1.0 and up because dropping it costs likeness. Check the LoRA's model page rather than reflexively lowering it.
• cpu_offload (auto / enable / disable, default disable) - offloads to system RAM to save VRAM at the cost of some speed. Leave it disabled unless you're tight on memory.

The output is a single patched MODEL that goes to your KSampler. Want more than one LoRA? That's the Stack nodes; this one is deliberately just the single case.

Installing it

Grab it through ComfyUI Manager (search the pack, install, restart), or clone it:

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

then restart. What the clone does not give you is Nunchaku itself. You need the official ComfyUI-nunchaku plugin installed and working first, plus a Nunchaku-quantized Qwen model downloaded. Nunchaku's own install is the fiddly part of the chain, it wants a prebuilt wheel matching your exact PyTorch and CUDA versions, so get plain Nunchaku Qwen generating before you add this. Python 3.11+.

Where people get burned

• ModuleNotFoundError: No module named 'nunchaku' means the base plugin isn't installed. Install ComfyUI-nunchaku and its wheel, then restart. This node is a passenger, not the engine.
• Your LoRA is a LoKr and nothing happens. A lot of newer character adapters ship as LoKr now, and this loader silently skips LoKR, LoHa and IA3 on Nunchaku models. If a LoRA has zero effect, check its format before assuming the node is broken.
• It's a development-status pack and chatty with debug logs. That's cosmetic and doesn't affect output; if the console spam bothers you, the logs are muted by default and only appear when you set nunchaku_log=1.