Nunchaku Z-Image-Turbo LoRA Stack V2

Read this part first, because it'll save you an afternoon: V2 only works with the unofficial Nunchaku Z-Image-Turbo DiT Loader. Specifically the one from ussoewwin's separate ComfyUI-nunchaku-unofficial-loader repo. It is not compatible with the official Nunchaku Z-Image-Turbo loader that ships in ComfyUI-Nunchaku. If you're on the official loader, this node will not do what you want, and V4 is the one you're looking for instead. So the single most important question with V2 is: which Z-Image loader are you running?

Assuming you're on the unofficial loader on purpose, here's what V2 is. It's a multi-LoRA stacker for Nunchaku-quantized Z-Image-Turbo with a dynamic UI, compatible with ComfyUI's Nodes 2.0 canvas. Z-Image-Turbo, for context, is Alibaba's 6B speed model that runs near-Flux quality on 12GB cards in around 30 seconds; quantized through Nunchaku's 4-bit SVDQuant it gets lighter still. As with everything in this pack, stock LoRA loaders can't patch the quantized weights, which is why these dedicated stackers exist.

How it works

You choose how many LoRA slots to show, fill each with a file and a strength, and the node applies them onto the model in sequence. Note that unlike the CLIP-carrying V1 and V4, this version outputs the model only, it's built around the specific way the unofficial loader exposes the model.

The inputs and outputs that matter

• model (MODEL) - your Z-Image-Turbo model, from the unofficial Nunchaku Z-Image DiT loader. This is the load-bearing requirement.
• lora_count (1–10, default 3) - how many LoRA rows appear.
• cpu_offload (auto / enable / disable, default disable) - offload to RAM for VRAM headroom.
• lora_name_1 … lora_name_10 and lora_strength_1 … lora_strength_10 - the per-slot file and strength (default 1.0). Set a name to None to skip a slot.

Output is a single patched MODEL for your sampler.

Installing it

Ships with the pack. ComfyUI Manager, search, install, restart, or clone:

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

then restart. Beyond the usual requirement of a working ComfyUI-nunchaku install (with its PyTorch/CUDA-matched wheel), V2 also needs the separate ComfyUI-nunchaku-unofficial-loader installed to provide the loader it's paired with. Python 3.11+.

Where people get burned

• The official-loader mix-up. This is the whole story with V2. Point it at the official Nunchaku Z-Image loader and it won't work; you need the unofficial one. If your outputs are wrong or the node errors on the model input, check which loader is feeding it before anything else, then switch to V4 if you'd rather stay on the official loader.
• Too many LoRAs on Z-Image. The community consensus is that Z-Image quality tends to collapse past one LoRA at a time. Keep it minimal.
• LoKr and refiner-layer LoRAs are skipped. Only Standard and PEFT LoRAs apply on Nunchaku; LoKR, LoHa, IA3 and context_refiner / noise_refiner layer LoRAs are silently dropped. A lot of Z-Image character LoRAs ship as LoKr, so this bites here. Check the format if a slot does nothing.