neonllama
This custom ComfyUI node transforms a core idea into a richly detailed positive prompt using a local [a/Ollama](https://ollama.com) LLM.
🧠 NeonLLama ComfyUI Extension
NeonLLama is a custom ComfyUI node that transforms one or more idea lines into vivid, richly detailed prompts using a local LLM — either through Ollama or LM Studio. It supports avoid content to guide generation and returns that list as the negative prompt.
🚀 Features
- 🧠 Converts simple ideas into rich, structured positive prompts.
- ✂️ Supports multi-line input — each idea line generates its own prompt.
- ⛔ Optional "avoid" list used to steer generation and returned as the negative prompt.
- 🔁 Retries with smart temperature adaptation to meet token targets.
- 🧮 Accurate token estimation using model-specific tokenizers.
- ⚙️ Fully configurable: model choice, token bounds, retry limits, and regeneration control.
- 🖨️ Logs all prompts and supports forced regeneration.
🧩 How It Works
-
Input one or more idea lines (line breaks separate them).
-
Optionally include avoid terms to bias generation away from undesired concepts.
-
The node:
- Attempts generation using your configured local model.
- Retries if the output is too short or too long.
- Logs inputs and outputs.
-
Outputs include the final positive prompt, your avoid list, and original idea.
📤 Outputs
| Output | Type | Description |
| ---------- | ------ | ------------------------------------------------------ |
| prompt | STRING | Combined positive prompt(s), separated with BREAK. |
| negative | STRING | Direct copy of avoid input, used as negative prompt. |
| idea | STRING | Original idea input, for UI traceability or debugging. |
📥 Inputs / Configuration Fields
| Name | Type | Description |
| ------------------- | -------- | ---------------------------------------------------------------- |
| model | Dropdown | Select the model name to use (fetched from the running backend). |
| idea | Text | One idea per line. Each generates a separate prompt. |
| negative | Text | Words to avoid — passed to model + used as negative prompt. |
| max_tokens | Int | Upper bound for output tokens (default: 75, max: 231). |
| min_tokens | Int | Lower bound for meaningful output (default: 50). |
| max_attempts | Int | How many times to retry if generation doesn’t meet criteria. |
| regen_on_each_use | Bool | Forces regeneration on every run, even with unchanged input. |
🧪 Example
Inputs:
idea:
haunted subway station with broken lights
avoid:
blood, gore, screaming
Outputs:
prompt:
dark abandoned subway, flickering fluorescent lights, cracked tiled walls, shadowy corners, old train cars, graffiti-covered pillars, dim green glow, debris scattered floor
negative:
blood, gore, screaming
🔁 Token-Aware Generation
- Token count estimated using model-specific tokenizer.
- Retries generation using adaptive temperature — starts hot and cools with each failure.
- Adjusts prompt by trimming or elaborating until token limits are satisfied.
- Prompts too short or too long are discarded unless they fall within the required range.
🔒 Robust Logging
- Logs prompts and failures to disk for traceability.
- Logs include selected model, prompt source, token counts, and generation time.
🧠 Smart Prompt Structuring
Prompts follow rules enforced via system instructions to the model:
- Use short, visually descriptive phrases, not sentences.
- Avoid storytelling, abstract descriptions, or emotion-driven language.
- Keep outputs grounded and aligned with original idea’s visual core.
- Connect elements with simple language: “with”, “under”, “surrounded by”, etc.
- Avoid repetition, overgeneralization, or verbose commentary.
🛠️ Requirements
-
At least one of:
-
ComfyUI installed and functional
-
Python packages:
requeststokenizerslmstudio(only required for LM Studio support)
📄 License
MIT License
❤️ Credits
Built by Neon Lightning ⚡ Powered by open-source local LLMs, punk energy, and rat affection 🐀