⚡| Style & Prompt Encoder

Z-Image Turbo is fast and photoreal, but if there's one honest complaint the community repeats, it's that raw Turbo is flat on style. The Style & Prompt Encoder (StylePromptEncoder2) is Martin Rizzo's answer, and it's the node that made his Amazing Z-Image Workflow famous: pick a look from a library of over a hundred hand-written styles - photographic, illustrative, wild, custom - and the node rewrites your prompt to match, then encodes it into conditioning. It's a style system, not a style LoRA, and it's the current stable workhorse of the whole pack.

Here's the thing that makes it work: Z-Image is prompted in natural language, and it responds to descriptive prompting. Each style in this library is a block of prose - "you are a cinematographer of the Golden Age of Hollywood…" - that gets woven around your subject. Your actual prompt stays intact; the style is the framing. That's a fundamentally different approach from the old Danbooru-tag style prompts of the SDXL era, and it's tuned for exactly how Z-Image (and its Qwen3-4B text encoder) thinks.

How it works

You give it a CLIP, your text, and a style from the dropdown (there's also a none option to encode unstyled). The node looks up the style template, applies it to your prompt (substituting your text where the {$@} placeholder sits), then tokenizes and encodes the result through your CLIP. The customization input lets you define or override styles using the >>> block format - same syntax as the pack's other encoders - so the built-in library is extensible without touching files. The gallery button opens the searchable, thumbnail preview of every style, which is the fastest way to shop for a look.

The inputs that matter

• clip (CLIP) - your text encoder; for Z-Image that's the Qwen3-4B model from LoadQwen34bModel.
• style (COMBO) - pick from the library or none.
• text (STRING, multiline) - your subject prompt.
• customization (optional) - your own >>> style definitions.
• Outputs: CONDITIONING (into your sampler) and STRING (the styled prompt, so you can see what was encoded).

Install

cd ComfyUI/custom_nodes
git clone https://github.com/martin-rizzo/ComfyUI-ZImagePowerNodes

Restart ComfyUI (or ComfyUI Manager → "Z-Image Power Nodes"). No pip dependencies. You need the Z-Image model triplet installed - diffusion model, Qwen3-4B text encoder, VAE.

Why this one over the G2.1 experimental?

The pack also ships Style & Prompt Encoder ^G2.1, which adds color palettes and fancier outputs but is explicitly experimental and uses a newer, thinner style definition set. This stable v2 node is where the mature 100+ style library lives, and it's the one the author recommends for day-to-day work. Where people get burned: expecting a style to override composition. Styles are an artistic filter on your prompt, not a control net - a style won't rescue a poorly-worded subject prompt. And if a style looks overdone, remember the library is opinionated on purpose; the customization field is there for you to tone one down or write your own. If you only add one node from this pack to your Z-Image setup, make it this one.