ZImage Turbo Single Prompt Conditioning

If you prompt Z-Image with one hand-written paragraph instead of four comma lists, this is the conditioning node you want. It's the single-prompt sibling of ZImageTurboConditioning: one multiline prompt box, one CLIP, and you get the same post-encode enhancement pipeline and the same three outputs - positive, negative, and the positive_prompt string.

The use case is LLM-written prompts. The pack actually ships a QWEN3VL_INSTRUCT.md that gives you a system prompt and a template for generating exactly the kind of 45-to-120-word single paragraph Z-Image likes - composition cues, material realism, a color palette, ending in masterpiece, high quality. You paste that output into this node and it just works. If that's your workflow, the four-field version is a chore you don't need.

Mechanically it's the same story as its sibling: the prompt is encoded with your CLIP (the Qwen3-4B text encoder), then the embedding goes through the "Capitan" basic stage - normalize → small MLP → blend back, with optional self-attention - followed by the advanced stage that sharpens detail with a tanh detail-boost and high-pass filter. All of that is re-implemented inside this pack, so there's no second repo to install.

The inputs you'll actually set:

• prompt - multiline, defaults to a portrait example. Write prose, keep it one paragraph. Z-Image follows natural language far better than tag lists.
• enhancement_profile - none, capitan_daily (default), or capitan_literal. Daily is the balanced profile; literal retains detail harder and can start looking crunchy on skin.
• seed - controls the random init of the enhancement MLP, so a fixed seed means a fixed enhancement. It does not touch the sampler seed.
• low_vram / device - low_vram forces fp16 and skips the self-attention step; device can pin the enhancement to CPU if VRAM is tight.

Outputs wire the same as the four-field node: positive → ZImageTurboSampler, negative is a zeroed conditioning (inert at CFG 1 - Z-Image Turbo is guidance-distilled, so don't bother writing negative prompts), and positive_prompt is the exact string you can save alongside the PNG.

Install is the same shared routine - ComfyUI Manager, search "ZImageTurboHQNodes", or:

cd ComfyUI/custom_nodes
git clone https://github.com/marduk191/ComfyUI-ZImageTurboHQNodes

then restart. Zero pip dependencies, and the model files aren't in the pack: you supply z_image_turbo_bf16.safetensors (diffusion_models), qwen_3_4b.safetensors (text_encoders), and a Z-Image VAE.

One honest caveat: the enhancement is applied on top of the raw CLIP embedding, so it won't rescue a bad prompt. If your one-liner is vague, the output is vague, just sharper. Give the LLM the template from QWEN3VL_INSTRUCT.md and keep capitan_daily on for most runs.