Ollama Prompt Refiner (Nukun)

This is the heavyweight of the pack and the reason a lot of people install it: a local Ollama model that takes your rough "word salad" prompt and rewrites it into a properly split prompt set for a specific base model - Pony v6, Illustrious, Pony v7, Z-Image, Anima, Wan 2.2 video, or Krea 2. No cloud, no API key, no censored frontier model deciding your subject is a problem. If an LLM-in-the-graph prompt enhancer is the thing you've been wanting, this is a well-built, extremely opinionated version of it.

The philosophy here is the right one, and it's the same one the community converged on: don't ask the LLM for one free-form rewrite, give it a tightly scoped job. The node asks for split outputs - base_prompt, foreground_prompt, background_prompt - plus a combined positive and negative, and it validates that the output is actually model-appropriate. For Pony v6 and Illustrious it pre-sorts your random words into fixed/base, foreground, background, style and discarded-noise buckets before the LLM ever sees them. That's the kind of guardrail a plain chat model won't give you.

How it works

You connect word_salad (typically from the pack's Random Vocab String List or Multi Vocab String List), pick target_profile, and the node talks to your local Ollama server at ollama_url (default http://127.0.0.1:11434) using ollama_model. The model dropdown refreshes live from Ollama's /api/tags, so pick the model, not the URL guesswork. Default model is autoren-darkidol-llama-3-1-8b:latest - a small abliterated Llama 3.1 8B, exactly the profile the KB says is right for this job: small, obedient, uncensored. German word_salad input gets translated to English first, which is a nice touch nobody else ships.

The pipeline is where it gets sophisticated. Default pipeline_mode = single is one request plus a JSON-repair retry when output is malformed. plan_compile adds a planner stage that first classifies your noisy input into a fixed JSON plan, then hands both plan and source to the compiler. plan_compile_review adds a semantic reviewer and at most one correction. Each mode costs more Ollama requests - two or three per run plus possible repair - so it's slower, but the review stage genuinely catches the "LLM invented a character you didn't ask for" failure that unconstrained enhancers are famous for.

The inputs worth setting as a beginner:

• style_anchor - fixed character names, LoRA triggers, quality tags that must survive the rewrite. Use it; it's the difference between "Ollama kept my character" and "Ollama invented a new one."
• target_profile - one profile per run. Pony v7 default.
• style_cluster (default 430) - the Pony v7 style_cluster_XXXX header value.
• prompt_mode - strict stays grounded in your text; creative lets it invent supporting detail and bumps sampling temperature.
• fallback_mode - adaptive rebuilds locally when Ollama fails, continue pushes through transport errors, strict stops with a detailed error. adaptive is the default and the sane one for unattended runs.

Outputs

positive and negative go to your text encoders. base_prompt/foreground_prompt/background_prompt are the split pieces for workflows that feed regional conditioning. report tells you what ran, plan_json and review_json show the planner/reviewer output (empty {} when that stage didn't run).

Installing it

Clone the pack, restart, and you're most of the way there - the node itself has no Python deps beyond the pack's own. The real dependency is Ollama itself:

cd ComfyUI/custom_nodes
git clone https://github.com/OnekoSL/Nukun_ComfyUI_Nodes.git

Then make sure Ollama is running (ollama serve), and pull a model: ollama pull autoren-darkidol-llama-3-1-8b:latest. The default context_length is 4096; the node unloads the model after each run by default to free VRAM for the diffusion pass - leave unload_after_run enabled unless you know why you need it off.

Common issues

Almost every problem is Ollama-side, not node-side. Malformed JSON gets one repair retry, then falls back per fallback_mode - if you see repeated fallbacks, your model is weak at structured output, so try a better one or lower temperature. If the model dropdown is empty, Ollama isn't reachable at ollama_url or /api/tags isn't answering. And a note for Pony v6/Illustrious: this node changed its output contract in newer versions - old workflows that expected separate pony_v6_* outputs need reconnecting to positive and negative. Read the report output when something looks off; it's unusually honest about what happened.