Prompts

Prompts is where you actually type in the old Searge workflow. It's the first link in a chain of input nodes, and it's the one with the text boxes - five of them, which is the giveaway that this pack had opinions about prompting.

It's from SeargeSDXL, the SDXL base-and-refiner workflow pack Searge released the moment SDXL 1.0 landed in 2023. Back then it was one of the go-to "everything wired up for you" setups, and the way it split a prompt into five parts was part of the appeal. The node lives in Searge/_deprecated_/UI/Inputs now - the v4.x rewrite moved to a data-stream architecture - but the five-field idea is still a decent lens on how SDXL prompting works.

How it works, and the five prompts

SDXL has two text encoders, and it responds to a "style" concept separately from the main subject. Searge exposed that as five boxes:

• main_prompt - the core subject/scene. This is your primary prompt.
• secondary_prompt - a supporting prompt, blended in depending on the prompt style you pick elsewhere.
• style_prompt - the look/aesthetic, kept separate so you can dial its influence.
• negative_prompt - what to avoid.
• negative_style - the style-side negative.

You don't have to fill all five. The point was to let you tune subject and style independently instead of cramming everything into one string - which is a real prompt-engineering technique, not Searge magic. How these five get combined is decided downstream by the prompt-style and prompt-processing nodes; this node just holds the text.

The inputs and outputs

The five prompts above are all STRING, multiline. Then there are optional connectors that make this the head of a chain:

• inputs (PARAMETER_INPUTS) - optional; lets you chain another input node before this one.
• image (IMAGE) and mask (MASK) - optional, for img2img/inpainting.
• Output: inputs (PARAMETER_INPUTS) - the accumulated settings bundle passed to the next input node.

That PARAMETER_INPUTS wire is the whole trick of the old Searge design: each input node adds its slice of settings to one growing bundle, and at the end a processor turns the bundle into usable parameters. Keeps the graph readable instead of a hundred loose wires.

How to install it

ComfyUI Manager: search SeargeSDXL, install, restart. Manual: cd ComfyUI/custom_nodes && git clone https://github.com/SeargeDP/SeargeSDXL.git, then restart. Manual installs need python -m pip install opencv-python in ComfyUI's Python environment first, per the README, or the nodes won't load. For actual generation you'll want the pack's models (SDXL base + refiner, fp16-fix VAE); the README lists them.

Common issues

If your style prompt seems to do nothing, that's expected until you set its influence in the prompt-processing node (style_prompt_power) and pick a prompt style that actually uses it - a couple of the styles ignore the style prompt entirely.

Bigger picture: this is a deprecated node inside a workflow built around the SDXL refiner, and the community has largely moved on from that pattern to single fine-tuned checkpoints. The five-prompt split is a nice teaching tool, but you don't need Searge to do it. And the eternal rule with this pack: load the latest workflow JSON with the latest node version, or you'll chase bugs that are really just version drift.