Nunchaku Qwen Image LoRA Stack V3

Of all the Qwen LoRA stackers in this pack, V3 is the one I'd reach for. It's the current design: per-row on/off toggles, a master toggle_all, dynamic UI that works on ComfyUI's Nodes 2.0 canvas, and the AWQ modulation fix always on with no switch to think about. If you're new to the pack and just want to stack a few LoRAs on a Nunchaku Qwen-Image model without fighting the interface, this is the default answer.

The setup it's built for: Nunchaku (MIT HAN Lab's SVDQuant engine) runs Alibaba's 20B Qwen-Image in 4-bit so it fits on consumer VRAM, and ordinary ComfyUI LoRA loaders can't patch those quantized weights. This pack is the fix, and V3 is its polished stacker.

How it works

You set how many LoRA rows you want, each row gets its own enable toggle, a file picker and a strength, and the node applies every enabled LoRA onto the model before it hits the sampler. The toggle_all control flips the whole stack at once, which is genuinely useful for a quick base-vs-LoRAs comparison. Where V3 pulls ahead of the older stackers is the AWQ modulation handling: Nunchaku AWQ-quantizes the model's modulation layers (img_mod.1 / txt_mod.1), and applying LoRA to them naively caused noise, so earlier versions gated the fix behind an experimental toggle. V3 just always applies it. One less dial, cleaner output.

The inputs and outputs that matter

• model (MODEL) - your Nunchaku Qwen-Image model from the DiT loader.
• lora_count (1–10, default 3) - how many LoRA rows appear.
• cpu_offload (auto / enable / disable, default disable) - offload to system RAM to save VRAM; off is fine unless you're memory-starved.
• toggle_all (default true) - the master switch across every row.
• enabled_1 … enabled_10, lora_name_1 … lora_name_10, lora_strength_1 … lora_strength_10 - the per-row toggle, file and strength. Default strength 1.0.

Output is one patched MODEL for your KSampler.

On strength: the reflex to sit everything at 0.7 is old SDXL advice. On modern bases people often run 1.0 and higher because dropping it loses likeness, so take the number from the LoRA's model page rather than a rule of thumb. And add LoRAs one at a time on a fixed seed, stacking is unpredictable and it's easier to spot which one misbehaves.

Installing it

ComfyUI Manager, search the pack, install, restart. Or clone:

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

then restart. It needs the official ComfyUI-nunchaku plugin installed and running first, with a Nunchaku Qwen model downloaded. Nunchaku's wheel must match your PyTorch and CUDA versions, which is the fiddly bit, so get plain Nunchaku Qwen generating before you layer this on. Python 3.11+.

Where people get burned

• A toggled-on LoRA that does nothing is usually a LoKr. This pack applies Standard and PEFT LoRAs only; LoKR, LoHa and IA3 are silently skipped on Nunchaku models. Check the file's format first.
• Lightning on lightning = black image. If your Nunchaku model already has a 4-step lightning distillation baked in, don't stack another lightning LoRA on top. If black frames appear without that, disable SageAttention and any torch.compile at startup, both are known black-image culprits with Nunchaku Qwen.
• It's a development-status pack. Expect frequent updates and verbose debug logging (muted unless you set nunchaku_log=1). The logs don't affect your image; they're the author debugging in the open.