CreaPrompt node 3

Node 3 is the rare CreaPrompt configuration with no humans in it. Its dropdowns are all scene, sky, style and camera - landscape, lighting, artistic style, artists, animals, image quality - which makes it the one you reach for when you're generating environments, concept art, or background plates instead of characters. If node 1 is the portrait machine, this is the establishing shot.

It reads the csv3 folder: Artistic_Style, Sky_Malapris_OCOMA, Image_Quality, Landscapes, Lighting, Artists, Animals, Places, and Cameras. Nine categories, all environment-flavored. The selection is clever about variety, too - Sky_Malapris_OCOMA is 77 lines of genuinely fun sky descriptions ("Apocalyptic Red: dust and smoke stain the sky the color of dried blood"), Artists has 3,600+ "style by..." entries, and Artistic_Style holds long, richly composed style paragraphs, some already carrying (weight:1.4) syntax for SDXL.

The inputs that matter

The recipe that works best here:

• Artistic_Style and Artists on 🎲random - this is the style engine. One gives you whole composed style descriptions, the other drops a named artist into the prompt.
• Landscapes, Places, Sky_Malapris_OCOMA on 🎲random - the subject and the backdrop.
• Lighting and Cameras on 🎲random - the look. The Cameras list is real cinema and still cameras ("ARRI ALEXA 65", "Canon R5"), and naming physical gear is high-leverage phrasing on modern models, not just an SDXL habit.
• Animals on 🎲random if you want a subject in the scene - 487 of them, from "a barking sea lion" to "a burrowing armadillo".
• Image_Quality - the quality-tag category ("masterpiece, highly detailed, 4k"). Keep it on if you're on an SDXL-lineage model; it's largely inert on LLM-encoded ones.

Optional inputs are the standard pack trio: Prompt_count (1–1000, emit N random prompts one per line), CreaPrompt_Collection (ignore categories, pull from collection.txt), and seed (reproducibility). Outputs are prompt (STRING) and seed (INT) - prompt into CLIP Text Encode, seed back into the KSampler.

Installing it

ComfyUI Manager, search "CreaPrompt", or by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/tritant/ComfyUI_CreaPrompt

Restart. No pip dependencies, no API key, fully offline. The author suggests the CreaPrompt_Ultimate checkpoint (CivitAI 383364) for best results - it's an SDXL Lightning/Hyper model meant for low CFG - but any checkpoint that likes descriptive prompts will take these.

Common issues

• New categories don't appear until you re-add the node or reload the workflow - dropdowns are built at node load. A CSV dropped into csv3/ shows up then.
• collection.txt is load-bearing. The pack's README says not to delete it and it means it; the node throws an error when it's missing.
• Tag-soup vs. prose mismatch. If you're running node 3 against a Flux or Z-Image checkpoint, the quality-tag and (weighted:1.4) categories read as literal text. Stick to the sentence-y categories - Landscapes, Places, Sky, Artistic_Style are natural language - and the output lands much better.
• Artist names are lottery tickets. 3,600 artists means plenty your checkpoint never saw. Not a bug, but don't be surprised when style by occasionally produces nothing recognizable.