🐸 LLM Prompt Refiner (Ollama)

There are two ways to get a fluent natural-language prompt for an image generator: hand-write it, or let a small local LLM do the translating. 🐸 LLM Prompt Refiner (Ollama) is the second - it takes a tag-based or mixed prompt ("danbooru-style tags" in the source's words) and rewrites it as one flowing sentence, through a model running on your own machine via Ollama. No cloud API, no key, no per-image cost.

The reason this class of node exists at all is worth understanding, because it's genuinely architectural. The KB's LLM-in-ComfyUI doc makes the point: if your checkpoint's text encoder is itself a language model reading an instruction - which is true of Anima (Qwen3 encoder), Z-Image, Klein, and others - then having a second language model write that instruction is translation between two things that speak the same language. The wildcards and comma-tag conventions that worked on SDXL don't always sing on an LLM-encoded model; a fluent sentence often reads better.

But the whole value hinges on one discipline: the LLM must not invent things. An unconstrained enhancer is a liability - it adds adjectives, props, and motivations you never asked for, and suddenly "1girl, red_hair" becomes "a mysterious girl with flowing crimson locks in a dark alley." The KB doc calls this subject drift, the number-two failure mode of prompt-enhancer nodes. The Refiner's system prompt is a wall of rules against exactly that: every word must be traceable to an input tag, unknown tags get passed through as plain text rather than guessed, colours must appear verbatim, hair must be written as "[length] [colour] hair", character names lead the sentence, and metadata tags like _(series) get dropped entirely. It's the rare LLM prompt you can audit.

Inputs and what they do:

• tags - the tag prompt to refine. Wire from a 🐸 Library or prompt node.
• model - Ollama model name, default qwen3:8b. Must be pulled first: ollama pull qwen3:8b.
• seed - same seed + same tags = same output (default 0).
• temperature - 0 is literal/consistent, higher is more creative; default 0.3.
• max_tokens - output length cap, default 220.
• prepend_tags - on by default, and the smart bit: it prepends the original tags before the natural-language prose, so your CLIP encoder gets both the precise tag signals and the fluent context. Turn it off if your encoder only wants prose.
• prepend / append - free-text slots for quality tags, LoRA triggers, or style cues to wrap around the output.

Outputs: prompt (the refined NL text) and debug (cleaned input, model settings, raw LLM output). Requires Ollama running locally at localhost:11434.

The positioning against its sibling is important and the source is explicit: this is a whole-scene rewrite node. 🐸 Tag to Description extracts appearance traits for character-builder workflows. Don't grab the wrong one for the job.

Installing it

Install the pack, then Ollama:

1. Pack: ComfyUI Manager → search Frog Node Pack → install, or cd ComfyUI/custom_nodes && git clone https://github.com/RabbitThatIsPink/FrogNodePack, restart, hard-refresh.
2. Ollama: install from ollama.com and make sure it's running (ollama serve).
3. Model: ollama pull qwen3:8b (or whatever you set in the model field).

No pip dependencies - the node talks to Ollama over plain HTTP (/api/generate), which is why there's nothing to install on the ComfyUI side.

Common issues

The big three: "connection refused"-style errors mean Ollama isn't running or isn't on localhost:11434. "model not found" means you set a model name that isn't pulled - run ollama pull <name>. And output that's not prose (chat preamble like "Here is your prompt:") is the classic dirty-output problem the KB warns about; the node's system prompt tries to suppress it, but if it leaks through, lower temperature and check your model - a big instruct-tuned model is more prone to chatting. Also: this node needs a local LLM that is comfortably uncensored for anime tags; if your model refuses, that's a model-choice problem, not a node bug - the community's standard answer is an abliterated or uncensored Qwen/Llama (the KB covers this exact pattern).