Generate Stable Diffsution Prompt With LLM

Yes, "Diffsution" is a typo baked right into the node's own name and output socket, not a typo in this article - worth knowing so you can find it in search and in your node list. Past that, this one does exactly what it says: you give it a rough topic, it asks a chat model to turn that into a detailed image-generation prompt, and hands you back a string ready to feed into your encoder. It's the "brainstorm buddy" node - useful when you know roughly what you want ("a lighthouse at night") but not the fifteen descriptive words that get a diffusion model to actually render it well.

How it works, and where it fits

Having an LLM write your image prompt for you isn't a novelty any more - it's become a mainstream step in a lot of ComfyUI workflows, and dedicated prompt-writing node packs draw real attention on release specifically because typing good prompts by hand is tedious. This node is a lightweight, DIY version of that idea baked into a general LLM-assistant pack rather than shipped as its own dedicated product.

The mechanism is the same template-substitution trick as its sibling nodes: template_system sets the model's role ("act as an assistant... don't over-interpret... answer in English"), and template_user is a longer, pre-written instruction that tells the model to act as an AI-painting prompt generator, pick from a list of aspects - lighting, background, art style, physical characteristics, behavior, demeanor, look, period, effects - and weave them into a description. Your prompt input drops into that instruction's {} placeholder as the topic. Edit template_user if you want a different style of output - tag-heavy for an anime checkpoint, prose-heavy for a photoreal one, JSON-structured for models that read it well.

One distinction worth being clear on: this is not the same thing as the LLM text encoders built into 2026-era checkpoints like Anima or Flux 2 Klein, where a language model is the encoder reading your prompt directly. This node runs before any of that - it's a separate helper step producing a plain text string. Whatever encoder your checkpoint actually uses reads this node's output exactly like it would read a prompt you typed by hand.

Inputs and outputs that matter

• prompt - your topic or rough idea, not a finished prompt. "A lighthouse at night" is a perfectly fine input.
• template_user - the pre-written prompt-generation instruction; edit it to bias the style of what comes back.
• model_name - a fixed dropdown of 21 chat models in provider/model-name form; the naming convention strongly suggests this pack calls OpenRouter, though the README doesn't name the provider directly.
• temperature (0–1, default 1) - worth pushing toward the higher end here specifically, since more variety is usually what you want from a prompt generator, unlike the Translate node where you'd want it low.

The output is a single string, unusually named stable diffsution prompt (typo and all) rather than the plain text its siblings use. Wire it straight into CLIPTextEncode.

Installing it

ComfyUI Manager: search comfyui-llm-assistant. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/longgui0318/comfyui-llm-assistant
pip install -r requirements.txt

Restart, then add a config.yaml with an API key - the README doesn't document the exact format, but given the OpenRouter-shaped model list, an OpenRouter key is the reasonable first thing to try.

Common issues

The default template_user explicitly tells the model to answer in English and to append a "Scene Introduction:" line after the tags - if you're generating prompts for a tag-based anime checkpoint, that trailing sentence is noise you'll want to strip or edit the template to drop, since most tag-trained models don't benefit from a prose sentence mixed into a tag list.

Same two structural gotchas as the rest of this pack: a missing or wrong config.yaml fails at run time, not at load time, so a node that imports cleanly can still error on every execution until the key is right; and model_name is a closed 21-item dropdown rather than free text, so if a specific model starts failing (several in the list are older free-tier names that get retired), swap to another entry rather than assuming your setup is broken.