Lackluster Prompt Enhancer

The name is accurate but understated: this isn't a generic "make my prompt longer" node. Lackluster Prompt Enhancer rewrites a rough idea into a structured generation or editing prompt using task-specific templates, and it expects to talk to a local LLM - Ollama by default, or any OpenAI-compatible endpoint. The task templates come from Bernini's official per-task prompt engineering, which is a nice pedigree: these are the shapes used for video editing, reference-driven generation, and the rest, not a one-size-fits-all "improve this."

How it works

You give it a prompt and pick a task_type - v2v (video-to-video), rv2v (reference-guided video editing), r2v (subject-driven video), t2v (text-to-video), t2i, r2i, i2i, i2v, mv2v, vi2v, ads2v, vrc2v, and more. It picks the matching system prompt and enhancement template, sends your text to the LLM at ollama_url (default http://127.0.0.1:11434), and returns:

• enhanced_prompt (STRING) - the rewritten prompt, ready for your text encoder
• negative_prompt (STRING) - a task-appropriate negative (or yours, if you set negative_prompt)

api_format chooses the wire protocol: Ollama (/api/chat) or OpenAI/vLLM (/v1/chat/completions). If ollama_model is left empty, the node passes the prompt through untouched - useful as a bypass while you wire things up. auto_enhance runs the enhancement on every queue; otherwise you trigger it with the Enhance button.

The multimodal trick: image0 through image4 are optional reference images, converted to base64 and sent to the LLM. Reference them in your prompt text as image0, image1, etc. - "make image0 wear a red coat" style instructions. That only does anything if your model is a VLM, so pick one that is if you're using this path.

The inputs worth caring about

• task_type - the most important choice; it decides the whole template
• ollama_url / ollama_model / api_format - your LLM plumbing
• temperature (0–2), max_tokens, seed - generation params. The tooltips carry a real caveat: GPT-5/o-series and OpenAI/vLLM endpoints ignore seed and may ignore temperature
• prepend_system_prompt - when True (default) it sticks the task's system prompt in front of the output, which matters because these models treat the output as an instruction - that's what CLIPTextEncode / T5 conditioning wants

Grounded expectations

The KB's take on local prompt enhancers applies here straight: this class of node removes the blank-page problem - it does not add image quality, and an unconstrained enhancer will happily add details you didn't ask for. Keep your instruction scoped and check the output before it hits a long queue. Also, "LLM as a Comfy node" has a security history - this node ships no such baggage, but it is arbitrary Python, so install it from the pack's repo via Manager like everything else.

Install and setup

Manager → "ComfyUI Lackluster Nodes" (requests + numpy deps). Then the real requirement is a running LLM server: Ollama installed and serving a model (ollama pull something in the 8B class is the community sweet spot), or a vLLM endpoint you point at. No LLM in the pack itself - the node is the client.