✨ Ollama Prompter

The "✨ Ollama Prompter" display name sells this node short, and slightly wrong. It can hand your keywords to a local Ollama model for rewriting - but with model_name on its default of "disabled" it's a completely self-contained prompt builder: nineteen dropdowns of art styles, real camera bodies, films, lighting rigs, and shot types that assemble an enhanced prompt without touching a single API. No Ollama, no key, no network. Start there; add Ollama later if you want the rewrite pass.

What it is

A prompt-enhancement node from ComfyUI Globetrotter Nodes (manifestations/comfyui-globetrotter). You give it keywords - the default is "a cat, sitting on a chair, fantasy art" - plus a stack of creative controls, and it returns a single enhanced_prompt STRING. In builder mode it's a much richer cousin of the pack's country nodes. With Ollama it becomes an LLM rewriter that takes your draft and expands it.

How it works

At load time the node reads style lists from data/styles/ (cameras, films, lighting, photographers, shot types, art styles) and prompt instructions from data/prompts/ - the cinematic, artistic-masterpiece, surreal-dreams, sdxl, and flux options come from those JSON files. It also pings Ollama's /api/tags endpoint and populates model_name with whatever models you have pulled.

When you run it, every dropdown set to "random" gets a value picked; then it assembles the prompt from your keywords, appending the chosen styles (camera: Nikon D850 with Nikkor 50mm f/1.8, film: Kodak Portra 400, etc.). If model_name is "disabled," that assembled prompt is returned as-is. If a model is selected, the prompt is sent to ollama_url (default http://127.0.0.1:11434/api/generate) with stream: false, your seed in the options, and a system prompt built from creative_mode, detail_scale, and the chosen prompt_style instructions. The code even strips <think>…</think> blocks out of the reply - a DeepSeek workaround that tells you exactly what people were running through this thing.

The inputs that matter

• keywords - the seed text. Everything else decorates it.
• model_name - "disabled" or a model you've pulled. This is the on/off switch for the LLM pass.
• prompt_style - random, disabled, or one of the five styles. Note random means a different style each run unless you pin it.
• camera, film, photographer, shot_type - the fun ones. Naming a real camera is genuinely high-leverage for photorealistic output, which is why this list is 49 bodies deep.
• creative_mode (standard/artistic/photographic/surreal/minimalist) and detail_scale tune how aggressively the LLM (or the builder) embellishes.
• seed - passed through to Ollama's options; pins the LLM's randomness.
• ollama_url - optional input, defaults to the local API. Change it if your Ollama runs elsewhere.

Install

cd /path/to/ComfyUI/custom_nodes
git clone https://github.com/manifestations/comfyui-globetrotter

Restart ComfyUI, or search "Globetrotter" in ComfyUI Manager. One gotcha: the node uses requests, which is not in the pack's requirements.txt (that lists jinja2, flake8, transformers). ComfyUI bundles requests, so it usually just works - but if you run a stripped virtual environment, pip install requests.

For the LLM pass you need Ollama running:

# outside ComfyUI - from ollama.com or your package manager
ollama pull llama3.2        # or whatever model you like
ollama serve                # keeps 127.0.0.1:11434 alive

Then restart ComfyUI so the node sees the new model in its dropdown.

Troubleshooting

• model_name only shows "disabled": Ollama isn't running or has no models pulled. Start it, ollama pull something, restart ComfyUI - the list is fetched when the node loads.
• The output literally starts with "ERROR: OLLAMA NOT REACHABLE": the node catches the failed request and returns that text as the prompt string. It won't stop your workflow - it'll happily encode the error into your conditioning. Check the output before wiring it downstream.
• Random results every run: prompt_style and the style dropdowns default to "random." Pin the ones you care about; seed only matters once a model is doing the rewriting.
• Timeouts: the request has a 60s cap. A big model on CPU can blow past that - try a smaller quantized model.