Z-Image Prompt

Let's get the name out of the way: Z-Image Prompt has nothing to do with Alibaba's Z-Image model. It doesn't load it, call it, or need it. The name borrows the vibe - Z-Image's community got famous for tightly structured, shot-by-shot prompting - but this node is just a text assembler that formats your prompt into labeled sections. You can point its output at Gemini, Z-Image, Flux, or a KSampler's text encode. It doesn't care.

It's the sibling of the pack's BananaPrompt, with a different template and a different default workflow. Where Banana Prompt is organized around medium-and-subject, Z-Image Prompt is organized around the shot itself: composition first, then subject, wardrobe, scene, lighting, mood, and a dedicated constraints slot for hard exclusions.

How it works

Same boring-good mechanism as its sibling. composition_or_framing is the only required field, and if you fill nothing else the node passes it through untouched - a prompt box in disguise. Fill any optional section and it assembles every non-empty section as [Title] followed by the body, blank-line separated. Note the brackets: the sections come out visually tagged, which reads cleanly to the LLM-encoder models that dominate 2026 prompting.

The sections

• composition_or_framing - required. Shot type, camera distance, angle, framing. The author's tooltip says it "usually defines the shot first," and it should: get the frame right before you worry about the subject.
• subject_or_identity - the main subject and its identity traits.
• wardrobe_or_appearance - clothing, hairstyle, makeup, visible accessories.
• environment_or_scene - the surrounding scene or background.
• lighting - light source, direction, softness, contrast.
• mood_or_style_or_quality - atmosphere, artistic intent, realism level.
• constraints - this is the one worth reading closely. Hard requirements and exclusions, per the tooltip: photorealism, no text, no watermark, avoiding unwanted artifacts.

That last slot is the reason to reach for this node over a plain text box. "No text," "no watermark," "photorealistic" - these are the failure modes people hit over and over with modern image models, and giving them their own labeled section instead of burying them in the body makes them dramatically more likely to stick. It's also a nice place to park negatives when the model you're using doesn't have a meaningful negative prompt.

Wiring it up

Single output: prompt (STRING). Feed it into BananaStudio's prompt input, a Z-Image prompt slot, or any other text consumer. It pairs especially well with the Gemini path because Gemini reads natural-language structure so well - the bracketed sections give it a clean schema to work from.

The one thing to know

Same caveat as Banana Prompt: it's a pure text formatter, so there are no failure modes to debug and no API key involved. The only way to be disappointed is to expect it to do the thinking for you. It won't enhance your prompt or expand variables (that's PromptEditor in this pack). What it does do - and does well - is force you to specify framing and lighting explicitly every time, which is exactly the habit that separates consistent prompters from people rolling dice.

Install

It ships inside tjcccc/comfyui_banana_studio, so:

cd ComfyUI/custom_nodes
git clone https://github.com/tjcccc/comfyui_banana_studio.git

Restart ComfyUI (or find "Banana Studio" in ComfyUI Manager). No models to download, no pip dependencies, no key. Requires ComfyUI ≥ 0.19.3.