Load Z-Image Turbo (SDNQ)

Z-Image Turbo is already the model that brought near-Flux quality to a 12GB card. This node asks the natural follow-up: what if you have less? Load Z-Image Turbo (SDNQ) loads a version of Z-Image Turbo that's been quantized down to 4-bit unsigned ints with an SVD rank-32 decomposition, so the whole pipeline sits in a much smaller footprint than the stock checkpoint. The author runs it at 2048x2048 on a 10GB RTX 3080 and reports roughly 7GB of VRAM used - and that he can keep an LLM running alongside it. That's the pitch: Turbo's speed and photorealism on hardware that normally has to pass.

The name is doing two jobs, so let's unpack it. Z-Image is Alibaba's 6-billion-parameter flow-matching model, and Turbo is the distilled 8-step variant that launched first and became the community's daily driver - the "guidance at CFG 1, 5-9 steps" workflow. SDNQ is the "SD.Next Quantization Engine," a PyTorch-native quantization stack from the SD.Next world that supports everything from 1-bit to 16-bit weights; the "-svd-r32" in the model ID means SVD rank 32. The default model_id is Disty0/Z-Image-Turbo-SDNQ-uint4-svd-r32, an Apache-2.0 repack that diffusers loads as a standard ZImagePipeline.

How it works, and what's different about it. This is a diffusers wrapper, not a ComfyUI-native loader. The node calls DiffusionPipeline.from_pretrained() directly with torch_dtype=bfloat16, applies whatever attention backend and offload options you asked for, and hands the ready pipeline back. Three practical consequences: the weights download into your HuggingFace cache on first run (not ComfyUI's models/ folder), ComfyUI's own memory management doesn't see them - the README warns you in so many words to "use carefully" - and it only works if your diffusers is new enough. You get a single ZIMAGE_SDNQ_PIPELINE output to wire into ZImageSDNQGenerate.

The inputs you'll actually touch:

• model_id - the HF repo. The default is the right one; you can point it at other SDNQ Z-Image quantizations, including Base versions.
• device - leave auto unless you have a reason not to.
• cpu_offload - sequential offloading when VRAM is tight. It works; it just costs speed.
• vae_tiling - tiles the VAE decode to save memory at high resolutions. Handy for those 2048px runs.
• attention_backend - default (SDPA) unless you're chasing speed. flash needs a prebuilt wheel on Windows; the author found sage was actually slower, and if you ever load a Base model, Sage is known to corrupt outputs entirely.

Installing. Search "ComfyUI-ZImageDit" in ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/erosDiffusion/ComfyUI-ZImageDit

Then restart ComfyUI. The README is admirably blunt that you'll probably need to pip-install things manually, and the big one is diffusers - it has to be recent enough for both the Z-Image pipeline and native SDNQ (≥0.40.0). From a portable install's embedded Python:

python.exe -m pip install git+https://github.com/huggingface/diffusers.git

The node also auto-installs the sdnq package if it's missing, falling back to a --no-build-isolation workaround for its known broken pyproject.toml; install_sdnq.bat is the Windows helper for that same fight.

Common gotchas. If startup fails, check the requirements - and note that quanto is not needed for these two nodes (it's for the pack's other, experimental files; ignore them). enable_compilation didn't work for the author; don't build a workflow around it. And the README owns up front that some options are only half-ported and untested. This is an alpha pack by a hobbyist who cheerfully admits he has no time to support it - it's genuinely useful, occasionally fiddly, and reads like it was written by someone who just wanted the model to work on his own 3080.