Z-Image Turbo Sampler

This is where the image actually gets made. Feed it the pipe from ZImageLoader, a prompt, a few dimensions, and it outputs a finished IMAGE - no separate KSampler, no VAE decode step bolted on afterward. That's the whole design philosophy of this pack in one node: instead of composing ComfyUI's usual sampler/decoder chain, it wraps Z-Image's diffusers pipeline as one call and does prompt encoding, denoising, and decoding internally.

What it's actually running

Z-Image-Turbo, from Alibaba's Tongyi-MAI lab, is a 6-billion-parameter distilled model - the fast, guidance-distilled sibling of the (much slower) Z-Image Base. Distillation collapses what would normally be 20-50 denoising steps into a handful, which is exactly why steps defaults so low and why classifier-free guidance mostly doesn't apply the way you're used to. If you've used any Turbo/Lightning/LCM-style model before, the mental model transfers: fewer steps, near-zero guidance, real quality ceiling, genuine speed. The README's own number is 24GB VRAM for a 2048x2048 image - fast for what it produces.

The inputs that matter

Six required, five optional. The ones you'll actually touch on day one:

• prompt - plain natural-language text, multiline. Z-Image's text encoder (Qwen3-4B) is bilingual, and the community has repeatedly found Chinese prompts get noticeably better adherence than English ones for the same request - worth knowing if the model keeps ignoring a specific detail.
• width / height - 256 to 4096, step 16, default 1024x1024. The node's ceiling is 4096, but don't chase it: Z-Image's useful native resolution tops out around 1-2 megapixels before quality degrades. Stay in 1024-2048 and upscale externally for more pixels.
• steps - default 4. Leaner than what the community settled on after actually testing this model - 8-9 steps is the commonly cited sweet spot, with diminishing or negative returns beyond that. If output looks soft or generic out of the box, bump steps to 8 before touching anything else.
• seed - default 42. Z-Image Turbo has a well-documented seed-sameness problem: different seeds with the same prompt tend to produce suspiciously similar compositions. Not a bug in this node - a known Turbo characteristic (Base fixes it, at a large speed cost). The community workaround is feeding an existing image in as a noise source at 0.55-0.75 denoise rather than relying on seed alone.
• guidance_scale - default 0, range 0-20. Guidance ships effectively off here, matching how Turbo was trained (it's guidance-distilled). That also means negative_prompt does essentially nothing at the default - CFG has to be doing something for a negative prompt to have leverage. Push toward 1.5-3 if you want it to matter; past ~3, images tend to overcook (oversaturated, "burned"), a trade-off the wider community has run into with this exact model.

The other three optional fields: batch_size (1-16, watch VRAM at high batch + high resolution), max_sequence_length (default 512, the Qwen3-4B token budget - raise it for long or structured prompts), and cfg_normalization/cfg_truncation, which - as their names suggest - exist to tame the guidance-burn problem above: rescaling and capping the signal so pushing guidance_scale past the model's native CFG-1 training regime doesn't blow the image out as hard.

The single output, IMAGE, is finished - wire it straight into a SaveImage or PreviewImage. There's no separate decode step to add.

How to install it

Same pack as ZImageLoader, so if you've already got the pack for the loader you've got this node too:

• ComfyUI Manager - search "ComfyUI-Z-Image-Turbo," install, restart.
• Manual - cd ComfyUI/custom_nodes && git clone https://github.com/tpc2233/ComfyUI-Z-Image-Turbo.git, then inside that folder: pip install modelscope, pip install git+https://github.com/huggingface/diffusers, pip install -r requirements.txt, restart.

This node is useless on its own - it needs a pipe from ZImageLoader upstream. There's no standalone model loading happening here.

Common issues & troubleshooting

Output looks flat or generic. Raise steps from the default 4 toward 8-9 first - that gap between the node's default and the community-tested sweet spot is real and it's the single biggest lever you have.

Out of memory at high batch size or resolution. No offloading toggle is exposed on this node, and the README's own VRAM figure is 24GB for a single 2048x2048 image - batch_size multiplies that. Drop batch size or resolution first if you hit OOM.

Type mismatch on pipe. Make sure it's coming from ZImageLoader specifically - this is a diffusers pipe object, not a standard ComfyUI MODEL, and it won't accept one.