Prompt Enhancer (Advanced)

Prompt Enhancer (Advanced) is the version of the simple Prompt Enhancer that lets you choose who the rewrite is for. The core trick is the prompt_style dropdown: instead of one generic "make this better" instruction, the node loads a researched system prompt tailored to a specific generation model. That matters because prompt formats aren't interchangeable - a prompt that sings on a CLIP-era SDXL is dead weight on an LLM-encoded Flux, and a video prompt needs temporal structure an image prompt never will.

The styles, and what they assume:

• generic - one polished generation prompt, no target assumptions.
• flux - natural-language FLUX prompt: subject, composition, materials, lighting, style, "return only the prompt."
• ltx_video - concise LTX video prompt: shot, subject action, camera motion, setting, lighting, temporal progression.
• wan - focused Wan video prompt: subject, action over time, environment, camera movement, composition.
• sdxl - concrete visual concepts, composition, lighting, viewpoint - the SDXL-appropriate shape.
• minimax_h3 and krea2 - load the pack's bundled skill files (in skills/), the same researched H3 and Krea 2 formats the simple node uses: H3's integrated_multimodal_description / soundscape fields and shot timestamps, Krea 2's one-rich-paragraph rule.
• custom - requires a system_prompt_override; otherwise the node raises "custom style requires system_prompt_override."

So the pairing is the point: generating video with LTX or Wan → pick the matching video style; generating images with Krea 2 → pick krea2; caption-free image work on Flux → flux. This is "have an LLM translate what you want into a well structured prompt following the guide" - the workflow the community actually converged on - with the guide already baked in per model.

Beyond the style, the inputs are the sampler controls the simple node hides: max_tokens (512 default, up to 4096), temperature (0.7), seed (0), and reasoning_effort (defaults to none). Turning reasoning off is deliberate - a prompt rewriter wants a model that follows a format and stops, not one that deliberates and leaks its scratch-work. The one output is enhanced_prompt, wired straight into your model's text encoder.

Honest expectations: these are the pack author's interpretations of each format, not the official prompt guides verbatim - read them as strong starting points. And a style-matched system prompt can't fix a bad model choice: a small quant will still rewrite thinly compared to the README's reference 27B Qwen abliterated model. If your idea is sparse, expect the enhancer to add detail; that's the feature, until it isn't - check the output before you feed it into a long generation run.

Install is the pack standard: ComfyUI Manager (search WepeNerd) or git clone https://github.com/WepeNerd/ComfyUI-WepeNerd.git into custom_nodes, pip install -r requirements.txt, restart. Model in ComfyUI/models/LLM, llama-server on PATH, and this node works - no projector needed for text-only enhancement.