Ollama Prompt Enhancer (Style Presets)

The class name says OllamaPromptEnhancerWithLoRA. There is no LoRA in it. No adapter, no weight file, no training - nothing. "LoRA" is aspiration (or marketing), which is why the honest display name is "Ollama Prompt Enhancer (Style Presets)." It's the sibling of the base Ollama Prompt Enhancer from the same pack, and its entire difference is that it writes the enhancement instructions for you, from a style menu, instead of making you type them.

What it actually is

You type a rough prompt, pick a style_focus and a detail_level, and the node assembles the LLM's instructions from two hardcoded dictionaries. Photorealistic gets "photographic details like camera settings, lighting, and realistic textures... professional photography terminology." Cinematic gets "camera angles, lighting mood, atmosphere, and film-like qualities." Artistic name-drops movements and artists. Anime gets a hand-wave about "aesthetic elements common in anime."

Then it runs the exact same pipeline as its sibling: POST /api/generate to your local Ollama at ollama_url, a temperature dial, and an unload of the model from VRAM after the call so it doesn't fight your sampler for memory. Outputs are enhanced_prompt and original_prompt, both STRING - wire enhanced_prompt into any CLIP Text Encode or text field.

The inputs that matter

Honestly, two of them. style_focus (general, photorealistic, artistic, cinematic, anime) and detail_level (minimal, moderate, detailed, very_detailed) are the whole point of this node. The rest - prompt, model, temperature, ollama_url - behave exactly like the base node. There are no optional inputs; max_tokens isn't exposed and is hardcoded to 256, so don't expect long rewrites.

When to reach for it

When you want a fast start and don't want to write instructions. The presets are decent prompt-engineering - naming equipment and photographic terms for the photo style matches the durable advice that "point-and-shoot film camera" moves the needle more than vague quality words. If you're doing LTX-style video prompts, you'll live in cinematic.

The trade-off is control: the preset owns the creative direction, and you can't nudge it the way you can with the base node's free-text enhancement_instructions. And the anime preset is the weak spot - it describes anime aesthetics in sentences, and if your target is a tag-trained anime checkpoint you'd want comma-separated booru tags, which this won't produce.

Installing it

Identical to the base node - it ships in the same pack. In ComfyUI Manager, search "Ollama Prompt Enhancer" and install, or:

cd ComfyUI/custom_nodes
git clone https://github.com/johnbonjohnsonn-creator/comfyui-ollama-prompt-enhancer

Restart ComfyUI. No pip step - the only import is requests, bundled with ComfyUI. The real dependency is a running Ollama server with a model pulled (ollama pull llama3.2), because this node is just an HTTP client for the /api/generate endpoint on http://localhost:11434 by default. No API key anywhere in this pack; that's the whole point.

Gotchas

Same two as its sibling, worth repeating because they're the confusing ones. First, silent failure: on any error - Ollama down, wrong model name, timeout - both outputs return your original prompt unchanged, so a dead server looks like the node "did nothing." The console lines starting [Ollama LoRA] are your only clue. Second, dirty output: the system prompt says "Return ONLY the enhanced prompt, no explanations or preamble," but that's a polite request, not an enforced constraint. A chatty model can still burp out preamble that lands in your conditioning. Check the first output before you trust the pipeline.