Style Prompt🧸

You type "a woman on a street." Style Prompt hands that to an LLM along with an art style you picked from a menu, and gets back a full, styled prompt: the descriptive elements, the lighting words, and a "style of [artist]" tag, all shaped to actually read as that style. That's the whole pitch. It's a prompt writer that knows more about art movements than you do, so you don't have to memorize what makes something look like Art Deco versus Chiaroscuro.

This is the original node in glibsonoran's Plush suite, and it's the one the pack was built around. The class name is Enhancer but everyone calls it Style Prompt.

How it works

Style Prompt doesn't generate an image. It generates text that a normal image model then generates from. Under the hood it sends the LLM a fairly detailed instruction telling it to build a prompt around your chosen style, weave in an artist or two who works in that style, and cap the length. The style menu (64 entries, from Anime to Biomorphic Abstraction to Photograph) is doing the heavy lifting, because the model's read on "what descriptive elements say Art Deco" is what makes the output land.

One wrinkle worth knowing up front: this node no longer picks the AI service itself. You wire in an AI Chooser node (also in Plush) that outputs an AI_Selection dictionary, and that's how you tell Style Prompt to use ChatGPT vs Groq vs Gemini vs Anthropic and which model. So Style Prompt on its own does nothing until that input is connected and you've got a valid API key in an environment variable.

The inputs that matter

• style (default Photograph): the heart of it. Pick from the 64-entry menu. This changes everything downstream.
• prompt_style (Tags or Narrative): this one's important and people get it backwards. Tags is the terse comma-list that Stable Diffusion and SDXL want. Narrative is the flowing-sentence form that DALL-E and older prompt encoders prefer. If you're feeding an SDXL checkpoint, use Tags. Our prompt-engineering notes back this up: comma-tag prompting is still alive and well on the whole SDXL lineage.
• artist (0 to 3): how many "style of" names to append. Set it to 1 or 2 and the style hits harder. Set it to 0 if you want to avoid naming anyone.
• creative_latitude (0.1 to 1.2, default 0.7): basically the LLM's temperature. Low sticks to your input, high improvises. Above ~0.9 it starts drifting.
• max_elements (3 to 25, default 10): caps how many descriptive chunks land in the prompt. SD weights the front of a prompt hardest, so more isn't better. Six to ten is the sweet spot.

Optional image input turns this into a vision node: feed it a reference image and the LLM describes that into a styled prompt. Optional prompt is your text seed.

Outputs: AI_prompt is the one you care about, wire it straight into your CLIP Text Encode. AI_instruction shows what got sent to the model (worth peeking at once). Style Info gives a paragraph on the style's history if you flip style_info on. Help and troubleshooting are text dumps for when things break.

Where it shines and where it doesn't

Style Prompt was designed and tested against SDXL base + refiner, and the author's right that it works best there. It's weak with turbo/LCM low-CFG workflows, because those don't fully render a long detailed prompt, and the whole point here is a detailed prompt. If your images look nothing like the style you picked, that's usually the cause.

Installing it

ComfyUI Manager is the easy path: open the Manager, Install Custom Nodes, search "Plush", install Plush-for-ComfyUI, restart. Manual install is cd ComfyUI/custom_nodes && git clone https://github.com/glibsonoran/Plush-for-ComfyUI.git, then cd Plush-for-ComfyUI && python -m pip install -r requirements.txt, then restart. The pack needs the OpenAI python library 1.3.5 or newer, which requirements.txt handles.

Common issues

The number-one gotcha is the API key. Plush reads keys from environment variables (OAI_KEY or OPENAI_API_KEY for ChatGPT, GROQ_API_KEY for Groq, and so on), and after you set one you usually have to reboot the machine before it's visible. If the node errors about a missing key, that's step one. The number-two gotcha is forgetting the AI Chooser: no AI_Selection connected, no output. And if you're on a hosted ComfyUI where you can't set your own env vars, the API-backed nodes just won't have a key to use. When anything's off, hook a text display node to the troubleshooting output; Plush logs a real reason there.