AI4ArtsEd Clean Prompt Randomizer

The AI4ArtsEd Clean Prompt Randomizer is an LLM that writes image prompts for you - but with a strict house style. Its system prompt bans camera references, film, lenses, depth of field, bokeh, motion blur, retro or vintage styling, film grain, and vignettes. The pitch is "media-neutral" description: contemporary scenes described as if observed directly by the human eye, sharp and natural, with rich detail about colors, lighting, textures, and atmosphere.

Why would you want that? Because the default habit of prompt generators (and the photo variant in this same pack) is to reach for photographic language - and that language carries an aesthetic with it. Portra means warm skin, Cinestill means night halation. The clean randomizer is the version that removes those crutches, which is genuinely useful when you're studying what a model does with raw visual description rather than with style keywords. Think of it as the control group for prompt experimentation.

How it works

A random_seed input exists purely to bust ComfyUI's cache: the node's IS_CHANGED returns NaN, so it re-executes on every queue pass, and the seed (which the code explicitly ignores) gives you a stable-looking thing to attach an Integer Randomizer to. Each run calls the selected LLM with the system prompt plus your optional context, and returns a fresh prompt.

The model dropdown is the interesting part. It lists ~30 OpenRouter models with per-model pricing and a capability tag (openrouter/google/gemini-2.5-flash [multilingual / $0.20/$2.50]), and it also auto-detects any models you have running in local Ollama and appends them as local/... entries. If an OpenRouter call fails, the code tries a smart fallback - same provider, then similar size, then a general small model. Local Ollama gets a similar fallback. That's a surprising amount of robustness for a pack this size.

The system_prompt is fully editable and multiline, so you can keep the mechanism and throw away its aesthetic opinions. That's the node's best feature.

The inputs that matter

• random_seed - INT, ignored functionally, needed to trigger re-runs.
• system_prompt - the editable prompt template. This is where you steer everything.
• model - OpenRouter or local Ollama model dropdown.
• context (optional) - a scene or concept the prompt should be built around.
• api_key (optional) - overrides the openrouter.key file.
• unload_model - "yes" frees the Ollama model from VRAM/RAM after generating.

Output: a single prompt STRING, ready for a CLIP/T5 encoder.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/joeriben/ai4artsed_comfyui_nodes
# restart ComfyUI

Or search "AI4ArtsEd" in ComfyUI Manager. For local models you also need the Ollama server running on port 11434; for OpenRouter you need a key (either in each node or in openrouter.key).

Common issues

"Empty model list." If the dropdown shows nothing, your ComfyUI restart didn't pick up the pack, or the model list failed to build - check that Ollama is reachable at localhost:11434 if you expect local entries.

LLM output is wordy or structured. The system prompt demands no meta-comments, but LLMs still occasionally emit titles or explanations. Edit the system prompt if it's not strict enough for you - that's what it's for.

It costs money. The default OpenRouter models are cheap (haiku-class at $0.80/$4.00 per M tokens) but not free. If you're generating a batch of fifty, switch to a local Ollama model or a -lite variant. The pricing is printed right in the dropdown, which is a nice touch you don't see in most packs.

One honest caveat: "media-neutral" is the goal, not a guarantee - a strong LLM can still sneak in photographic language despite the rules. It's an experimental pack from an academic project, and the output quality varies with model choice. For a paid hosted model it's surprisingly good; for a 3B local model, temper expectations.