Nunchaku Z-Image DiT Loader

Z-Image is already the small, fast one - 6B parameters, ~30 seconds an image on a 3060. So why would you 4-bit it? Because "already fast" and "faster" aren't the same thing, and on a tight card the VRAM headroom is real. NunchakuZImageDiTLoader loads a Nunchaku 4-bit build of Z-Image and hands you a MODEL, and Nunchaku's own v1.2.0 notes claim a 20–30% Z-Image speedup on top of everything the model already does well.

Z-Image earned its spot the honest way: it landed in Flux 2's launch window, ran on the RTX 3060 12GB that most of the community actually owns, and did it uncensored under Apache 2.0. The community started calling it "SDXL 2.0." Putting it on Nunchaku is for people who want that same daily-driver model to sip even less memory or turn out frames quicker.

How it works

Same SVDQuant engine as the rest of the pack: outliers get peeled into a small high-precision low-rank branch, the bulk runs in 4-bit, and Nunchaku's fused kernel keeps the low-rank part from costing speed. The output stays close to the full model while the footprint drops hard. On a model this size the practical effect is less "now it fits" and more "now it's quick and leaves room for the encoder, LoRAs, and a big batch."

And the usual rule holds: this loads a Nunchaku-quantized Z-Image file, not the stock Tongyi-MAI weights and not a GGUF. Grab the build from Nunchaku's HuggingFace.

The inputs and outputs that matter

This one's refreshingly simple - a single required input:

• model_name - the pre-quantized Z-Image file. Empty dropdown means you haven't downloaded one into your models folder yet.

Output is a MODEL that wires into your KSampler like any Z-Image graph. Remember Z-Image Turbo's settings: it's guidance-distilled, so run it at low steps (around 8–9) and CFG 1, and negative prompts do nothing at CFG 1 - that's the model's nature, not the loader's.

How to install it

Part of the ComfyUI-nunchaku pack. ComfyUI Manager → search "ComfyUI-nunchaku" → install, or

cd ComfyUI/custom_nodes
git clone https://github.com/mit-han-lab/ComfyUI-nunchaku

then restart. Then the two dependencies this pack always needs: the backend wheel (install with NunchakuWheelInstaller) and a quantized Z-Image model from Nunchaku's HuggingFace org - INT4 for RTX 20/30/40-series, NVFP4 for Blackwell 50-series - dropped in models/diffusion_models.

Common issues & troubleshooting

Dropdown's empty. No quantized Z-Image file present. Nunchaku won't take a normal Z-Image safetensors; download their build.

Node won't load / import error. The backend wheel isn't installed or doesn't match your Torch. Run NunchakuWheelInstaller, restart.

Only recent Nunchaku versions have Z-Image. Z-Image support arrived in the v1.1.0/v1.2.0 wave (late 2025 into early 2026). If your NunchakuZImageDiTLoader is missing or won't load a Z-Image file, update the node pack and the engine wheel - an old wheel won't know the format.

Seeds all look the same. That's Z-Image Turbo, not Nunchaku - the distilled Turbo has notoriously low seed variance. If you want real variety, that's a case for the non-distilled Z-Image Base (run at higher steps and CFG 3–5), not something the loader changes.