RPG Art Style Selector

RPGArtStyleSelector is the style half of the ComfyUI-RPG-Characters pair. RPGCharacterSelector tells the model who your character is; this node tells it how to draw them. One dropdown, six styles - Anime Style, Dark Fantasy, Realistic, Fantasy Illustration, Digital Painting, and the newer Sci-Fi / Cyberpunk - and it spits out four text outputs: a short positive prompt, a short negative prompt, and two long LLM-facing descriptions.

You could fake this with a text file of style tags and a bit of discipline. The value is that it's already written, it's consistent, and it's wired for both of the pack's pipelines - standard and Ollama - without you having to maintain a second prompt bank.

How it works

Under the hood it's two hardcoded dictionaries. STYLE_PROMPTS holds the short, tag-style positive/negative pairs (e.g. Realistic → "photorealistic, natural lighting, high detail, sharp focus"). A second map holds the verbose, prose-style descriptions - the "1024x1024 Extreme Close-Up Portrait" essays that walk through eyes, hair, clothing, and background for each style.

So the same dropdown drives two very different outputs aimed at two different consumers: the tag pairs go to the standard CLIP encoding path, and the prose goes to an LLM.

The outputs, and the trap in the fourth one

• positive_prompt / negative_prompt - short tag strings, ready to feed the pack's Prompt Concatenator or a plain CLIP Text Encode.
• Ollama_Posative_Textbox_1 - the long style description, meant for ComfyUI-Ollama's OllamaGenerateV2 node (yes, "Posative" is misspelled in the output name; that's the author's own label, don't fight it).
• Ollama_Negative_Textbox_1 - here's the trap. Despite the name, this is not a negative prompt. It's an instruction template telling an LLM to write a negative prompt for you: "You are an AI art generator… identify unwanted visual artifacts…" Wire this straight into negative conditioning and you'll get prose in your negative, which is worse than no negative at all. It belongs on the text input of an LLM node, and whatever the LLM returns goes into conditioning.

That's the classic mix-up with this node, and it's the main reason people bounce off the Ollama path.

Install

Same as the rest of the pack - ComfyUI Manager, search RPG-Characters, or:

cd ComfyUI/custom_nodes
git clone https://github.com/lord-lethris/ComfyUI-RPG-Characters.git

Restart ComfyUI. No requirements.txt, no model files, no extra dependencies for the string outputs.

Common issues

• The negative instruction isn't a negative. Read the output before wiring it anywhere. It's an LLM instruction, not a prompt.
• Nothing on the Ollama outputs. They only do something if you've installed stavsap/ComfyUI-Ollama and have an Ollama server running. The README says exactly this; without both, those outputs are inert strings.
• One tip on the Ollama path: running an LLM next to a diffusion model can starve your VRAM. If your sampler suddenly crawls after enabling the Ollama branch, that's your local LLM hoarding the card - worth knowing before you blame the node.

The short-tag outputs are the dependable ones. If you just want to stop writing "dark fantasy, moody lighting" from memory every time you change style, this node is a fine little memory.