toobusy Z-Image Turbo

Z-Image Turbo is Alibaba's Apache 2.0 model that took the local photorealism default by late 2025 - it's fast on 12GB cards, and realism families like Juggernaut and epiCRealism rebuilt their lines on it. But a full Z-Image Turbo graph in vanilla ComfyUI is loaders plus LoRA plus text-encode plus sampler plus decode, strung across half the canvas. toobusy Z-Image Turbo folds all of that into one node: model, CLIP and VAE loaders, up to five LoRA slots, conditioning, sampling, and decoding, with the whole chain inside.

The README's before/after table says it plainly: "UNET/CLIP/VAE loader + LoRA + encoding + sampler + decode → 1 node." And because it's a fold rather than a wrapper around an API, every sub-piece is still accessible - the node passes the loaded model, clip, vae, and the final positive/negative conditioning out of its outputs so you can reuse them downstream without re-loading anything.

How it works, and the inputs that matter

Model auto-detection: since v0.2.10 the node does a fuzzy filename scan of your model folders, so it usually finds the right Z-Image diffusion model, lumina2 text encoder, and VAE on its own. If it grabs the wrong file, set model_name / clip_name / vae_name explicitly. The default VAE is FLUX1/ae.safetensors - that's not a mistake, Z-Image shares Flux's VAE.

The three fields you'll touch most:

• positive / negative - plain text; the node encodes them with the internal lumina2 CLIP.
• ratio_preset + megapixels - the resolution system, 2:3 at 1MP by default. Set both width and height to positive values to override.
• steps, cfg, denoise - the sampling knobs. Defaults are 8 steps, cfg 1, denoise 1. Z-Image's LoRA-stacking ceiling is real (about two LoRAs at strength ≤1.4 before things melt), so don't expect to pile on five at full strength and keep the photorealism.

The outputs are the story

Beyond image, it exposes latent, width, height, and - the point for power users - model, model_clean, clip, vae, positive, negative. That's the whole pipeline out for the picking. The README's recommended hires loop is:

Z-Image Turbo -> toobusy Hires Upscale -> Z-Image Turbo (latent_override)

Where people get burned: model vs model_clean. If you're continuing at the same resolution and same control conditions, model is fine. If you're doing a second pass at a different resolution or swapping LoRAs, use model_clean - the README calls it "safer" and it is. Also note latent_override beats both the image input and the empty latent when connected.

Connect an image and the node switches to img2img (VAE-encoded start latent, denoise becomes change strength). Connect a zit_control from toobusy ZIT ControlNet and you get depth/canny/pose control. Leave both unconnected and it's plain text-to-image. That tri-mode behavior is why this node collapses a whole graph - it's not a button, it's a pipeline with the plumbing exposed.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/nicekriss/toobusy.git toobusy

or search "toobusy" in ComfyUI Manager, then restart and hard-refresh. You supply the model files yourself: ZIT/zImage_turbo.safetensors, the lumina2 text encoder, and the Flux VAE in your model folders. The example workflow is docs/workflows/z_image_turbo.json.

If you already have a working Z-Image Turbo graph you love, this node is optional - you're not gaining quality, you're gaining a shorter canvas and a reusable one-node pipeline. If you're setting up Z-Image from scratch, it removes most of the setup pain in one drop.