Z-Image Turbo (TJ)

A distilled-model one-shot: loader, encoder, sampler, and LoRA stack in one node

Z-Image Turbo (TJ) is a self-contained generation node for the Z-Image Turbo model: it loads the diffusion model, the Qwen3 text encoder, and the VAE; encodes your prompt; applies up to five LoRAs; optionally runs ZIT ControlNet; samples; decodes; and hands you the image - plus the model, latent, and conditioning for whatever you want to do next. It's the "everything from prompt to pixels" node, Z-Image edition.

Why a single node for this? Because Z-Image Turbo is a distilled, few-step model (the defaults: 8 steps, cfg 1.0, euler/simple - the classic turbo recipe where CFG stays low and steps stay short), and for a fast iterative model, the friction of wiring the standard loader→CLIP→sampler graph for every test kills the whole point. This collapses that graph into one node, and its setnode_name defaults to "Z-Image" because it's built to also be a wireless provider.

How it works

The node wraps the standard Z-Image Turbo graph. ratio_preset + megapixels + divisible_by compute your width/height (leave width/height at 0 and it does the math), and aura_shift (default 3) is the Z-Image-specific sampling tweak. The lora_1_enable...lora_5_strength slots stack up to five LoRAs inline - no separate LoRA loader nodes cluttering the graph.

The overrides are where it gets flexible: image flips it to img2img, latent_override wins over image/empty-latent, positive_override/negative_override replace the text with ready-made conditioning, and zit_control accepts the bundle from ZIT ControlNet (TJ) for depth/canny/pose control. You can also bypass the built-in loaders entirely with model_override/clip_override/vae_override.

The outputs that matter

The headline one is image. But it also passes through latent, model, clip, vae, and positive/negative conditioning, plus width/height - so you can chain a second pass (upscale, img2img refinement) without re-running the whole pipeline, and model_clean gives you the un-LoRA'd model for comparison. That output breadth is the difference between "a generator" and "a pipeline component."

Install it and get the models

Install the pack, then place the models exactly where the node expects them:

cd ComfyUI/custom_nodes
git clone https://github.com/designloves2/ComfyUI-TJ_NODE.git

Model paths from the README: models/diffusion_models/ZIT/z_Image_turbo_bf16.safetensors, models/text_encoders/Qwen3/qwen_3_4b.safetensors, models/vae/z-image-Vae.safetensors, and optionally models/loras/Lora/lina/LINA_ZIT.safetensors. Drop the model files in those spots, restart, and the dropdowns populate.

Where people get burned

The settings are the trap for people coming from full-scale models: Z-Image Turbo is a turbo, so cranking cfg to 7 and steps to 30 is actively wrong - keep them low (the defaults are the point) and tune aura_shift instead if output feels off. Second, the LoRA slots are per-slot toggles, not auto-applied: enable lora_N_enable or you'll wonder why your LoRA did nothing. And if you connect zit_control but nothing changes, check that ZIT ControlNet actually has a condition enabled - an empty bundle is a no-op. On an 8GB card this model is genuinely usable because it's distilled and fast - that's the whole reason this one-node wrap exists.