TwinFlow Z-Image-Turbo

Here's the thing most people miss about this node: Z-Image is the model the local realism community migrated to in late 2025. When everyone got tired of SDXL's limits and FLUX's VRAM appetite, Z-Image Turbo became the default for photorealistic work - it's the base under a huge chunk of the newest realism fine-tunes. This node is Z-Image Turbo running on Runware's servers, so you get that modern look without owning the GPU that usually demands.

The "Turbo" part is a distillation story, and it changes how you set it up. The Turbo weights collapse the diffusion trajectory into a handful of steps - that's why the node's steps defaults to 4 and CFGScale to 0. This is not a bug or a stripped-down preset; it's the whole point. The distilled model doesn't want 30 steps and guidance of 7. Feed it the modern recipe and it's fast, and quality-wise it punches far above what "4 steps" sounded like in the SDXL era.

Inputs that matter

positivePrompt, width, height are required (1024×1024 default, 16px step - Z-Image likes multiples of 16). Then:

• steps - default 4, range 1–50. The distilled default is right; only climb if you're chasing fine detail and willing to pay for it.
• CFGScale - default 0. On a guidance-distilled model, raising CFG mostly buys prompt adherence at the cost of quality. If you must nudge it, stay under ~2. This is the knob people instinctively crank and then complain the output is "burnt."
• scheduler - 76 options, leave on (default).
• lora - a RUNWARE_LORA socket, which is the killer feature here. Z-Image's whole local ecosystem is its LoRA library; on the hosted side you stack LoRAs via the pack's Runware LoRA builder the same way you would locally.
• negativePrompt - remember the caveat: ignored when guidance is off (CFG ≤ 1). Since this node defaults to CFG 0, your negative prompt is doing nothing unless you raise guidance.
• acceleration / acceleratorOptions - hosted-side optimization levels (none/low/medium/high) and the accelerator builder. "High" shaves latency; the trade is usually quality or consistency.
• numberResults - up to 20 variants per run.

seed, outputFormat, outputQuality, and the safety/ttl cluster are standard pack furniture.

How it works

The node sends runware:twinflow-z-image-turbo@0 as an imageInference task through the Runware SDK. The image output comes back as a native IMAGE tensor - wire it to Preview Image or Save Image, or feed it into other Runware nodes as a seedImage. The title bar shows the per-run cost, which is useful because this model is fast enough that you'll stop noticing you're spending money.

Install and gotchas

cd ComfyUI/custom_nodes
git clone https://github.com/Runware/ComfyUI-Runware
pip install -r ComfyUI-Runware/requirements.txt

Restart, paste your Runware API key into Settings (or RUNWARE_API_KEY), and you're live. Paid, per image, minimum top-up.

Where people trip: they import SDXL habits - 30 steps, CFG 7, negative prompt as a crutch - and get results that look worse than the defaults. Trust the Turbo defaults, keep CFG near zero, and if a generation looks flat, adjust the prompt before you adjust the sliders. And don't expect offline use: this is a hosted model. If you want Z-Image locally, you're looking at a different, much heavier install - this node is the "no-GPU edition."