Text-to-Image Prompt Enhancer

The most common reason a generation looks wrong isn't the model - it's that you wrote the prompt in the wrong dialect. Illustrious wants Danbooru tags, Flux wants a natural-language paragraph, and most of us only speak one of those. This is the node from Eric's Prompt Enhancers pack that handles that translation for you. Type "portrait of a warrior with red armor at golden hour," pick your target platform, and it returns a prompt written the way that model actually likes it - plus a matching negative.

It's the most powerful and the most fiddly node in the pack, and it earns its place. The community's own advice for LLM-encoded models is basically "have an LLM translate what you want into a well-structured prompt following the model's guide," and that's exactly what this does - on a timer, in your graph, with no copy-pasting.

How it works

Under the hood it's the same local-LLM call as the pack's other nodes - no built-in model, talks to LM Studio or Ollama over an OpenAI-compatible endpoint (default http://localhost:1234/v1). The difference is the knowledge baked into its system prompts: 14 target platforms (flux, flux_kontex, sd_xl, sd_1_5, pony, illustrious, chroma, pixart_sigma, aura_flow, noobai, kolors, qwen_image, and more), each with its own prompt style, optimal token length, and negative-prompt conventions. Pick pony and you get score tags and a comma list. Pick flux and you get a clean paragraph - because a T5 encoder treats a tag soup as noise.

It also understands the pack's syntax: (keyword:1.5) for emphasis and {a|b|c} for alternation, both passed through to the LLM and preserved. And it can caption optional reference images (up to two IMAGE inputs) with a vision pass, guided by a directive like recreate, style only, or subject only.

The inputs that matter

There are a lot of them. Here's what you actually touch:

• target_platform - the whole point of the node. Set it to the model you're about to run.
• text_prompt - your idea. Or, via prompt_context, an instruction: expand_short_prompt, finish_opening_line, or prompt_from_item_list turn a list of elements into a scene.
• creative_randomness - off/none/suble/subtle/moderate/bold/storyteller/chaotic. none barely embellishes; storyteller writes you a mini-narrative. This is the fastest way to get more interesting (or more unhinged) output.
• genre_style - 20+ options from cinematic to horror to cyberpunk. This and target_platform cover 90% of use.
• positive_keywords - where your LoRA trigger words go. If a style LoRA needs its trigger in the prompt, put it here so it survives the rewrite.

Everything else - camera angle, composition, lighting source/quality, weather, historical period, framing, pose - defaults to none and only matters when you specifically want to steer something. They're a real trap for over-controlling: set a half-dozen and the LLM stops being creative and starts ticking boxes.

A word about that default negative

The node ships negative_keywords pre-filled with the classic forty-word block (ugly, blurry, bad anatomy, extra fingers...). Here's the thing: on modern distilled models that run at CFG 1, negative prompts are computed at weight zero - they literally do nothing, and the KB's corpus research shows the long boilerplate block reads as cargo even where negatives do work. On Pony/Illustrious it's fine and the pack also auto-generates platform negatives. On Flux or any distilled checkpoint, you're often better off deleting it and restating what you want in the positive.

Outputs and install

Wire positive_prompt into your CLIP Text Encode or conditioning, negative_prompt beside it. vision_caption shows the caption it wrote from your reference image, settings_used summarizes what it applied, seed_used tells you the seed behind the generation when seed_mode is fixed or increment (handy for batched variations). status covers the boring stuff.

Install is shared with the pack: ComfyUI Manager → search "Local LLM Prompt Enhancer", or git clone https://github.com/EricRollei/Local_LLM_Prompt_Enhancer.git into custom_nodes, pip install -r requirements.txt, restart. The hard requirement is just requests; the heavy transformers/accelerate/bitsandbytes stack is only for the optional local Qwen3-VL vision backend. Vision captioning works without it too - point vision_backend at a vision-capable model loaded in LM Studio or Ollama.

One more thing worth knowing: this pack is dual-licensed (non-commercial CC BY-NC plus a commercial license), so if you're building a product on it, check the license before you ship.