🦙 Bulk Prompt Ollama Improver

Wire a Bulk Prompt Loader into this node and every prompt in your batch gets rewritten by a local Ollama model before it ever reaches a CLIP Text Encode. That's the whole job, and it's a good one: you feed it a rough idea like "a red apple on a wooden table" and it hands back a more vivid, detailed version, offline, uncensored, free per call. The reason to care about this node specifically is that it solves the one problem that sinks every other prompt-enhancer: chat models love to wrap their answers in "Sure! Here's an improved version:" and markdown fences, and that scaffolding lands in your conditioning as literal prompt tokens if nobody strips it.

How it works

It uses the official ollama Python library against a running Ollama server - so Ollama is a separate process, not something loaded in-graph. Each run calls generate() with your prompt, your instruction as the system prompt, and a keep_alive value. The node is version-tolerant across ollama-python 0.4→0.6 and declares VALIDATE_INPUTS, so the dynamically-fetched model list passes ComfyUI's backend validation instead of throwing the classic "unknown required inputs" error on load.

The star is trim_output (on by default). After the model replies, the node runs a real cleaning pass: leading preambles, --- rules, ``` code fences, Prompt: headers, surrounding quotes, and trailing "let me know if you'd like changes!" offers all get stripped, leaving just the prompt. The KB's writeup on local-LLM prompting calls dirty output the #1 failure mode of this category of node - a model's conversational habits bleeding into your conditioning is worse than no enhancer at all. This node bakes the fix in. Turn trim_output off and you get the raw reply if you want to see what the model really said.

Inputs and outputs

Set these and forget them:

• url - default http://127.0.0.1:11434. Change it for a remote host (http://192.168.1.50:11434), then hit 🔄 Reconnect to refresh the model dropdown.
• model - populated by clicking Reconnect. Pick the one you ollama pull'd.
• instruction - the system prompt telling the model what to do. The default ("Rewrite the following prompt to be more vivid and detailed. Output ONLY the improved prompt...") is genuinely good; tighten it if you want a specific style.
• timeout - default 120 seconds. Bump to 300+ for big models (14B+), because the first call also has to load the model into VRAM and that counts against the budget.
• keep_alive_minutes - how long Ollama keeps the model resident after a call. -1 keeps it forever, 0 unloads immediately. If you're batch-running hundreds of prompts, this is the difference between fast and "waiting for a model load every row."
• enabled - flip off to pass prompts through unchanged, no Ollama call. Cheap A/B toggle.

Optional context input accepts comma-separated ints for multi-turn continuity. Outputs: result (the improved, trimmed prompt - wire it to your CLIP Text Encode), context (Ollama's context tokens, for chaining), and meta (JSON with the model, timing, token counts, and a trimmed flag - handy if you want to see how long the rewrite actually took).

Installing it

The one node in the pack with a dependency. The pack's requirements.txt pins ollama>=0.4.0; ComfyUI Manager will install it automatically, or do it yourself:

cd ComfyUI/custom_nodes
git clone https://github.com/sifatrads/comfyui-bulk-prompt.git
pip install ollama

Then make sure Ollama is running (ollama serve) and pull a model, e.g. ollama pull qwen2.5:7b or llama3.1. The pack's __init__.py is defensive here: if the ollama package is missing, this node is silently disabled but the core loader nodes keep working - so a missing import won't take down your whole batch setup.

Where people get burned

• Timeouts on the first few rows - that's the model loading into VRAM. Raise timeout, or set keep_alive_minutes to -1 so it stays warm.
• Model dropdown is empty / "No Ollama model selected." - click 🔄 Reconnect. The list is cached; pulling a new model in the terminal won't show up until you refresh.
• Ollama on a different machine - you can point url anywhere reachable, and the node normalizes a bare localhost:11434 to a proper URL for you.
• VRAM - you're now budgeting for two models at once (LLM + checkpoint). Keep the LLM small (an 8B, or 3B on tight cards) and lean on keep_alive so it doesn't sit resident when you're not batching. The KB's guidance is blunt: the job doesn't need a big reasoning model, it needs a small obedient one that follows a format and stops.