LLM 文本生成器

This is the node that makes the whole pack make sense. LLMTextGenerator runs a local language model inside your ComfyUI graph and hands you the generated text as a STRING - no API, no key, no Ollama server in the loop. Type a rough idea, it writes you a structured prompt. Feed it an image, it captions it. Wire its output into a text socket, and your prompt becomes whatever the model decided, on every run. It's the "LLM as a tool" pattern - the thing the KB tracks as the most common LLM-in-the-graph job: turn a blank-page idea into a proper prompt before the sampler ever sees it.

The node is the tail end of a two-node pair. Its model_instance input comes from LLMModelLoader (also in this pack, under "LLM 模型加载器"), which is where you pick the model folder and choose text or multimodal mode. Text mode for Gemma-2/Llama/Qwen-class text models; multimodal for Qwen2.5-VL, Gemma3, and other vision-capable ones - the README is explicit about this, and picking wrong is the most common mistake: in text mode, any image content is silently dropped with a warning.

How it works

It builds a chat message list, applies the model's own chat template (apply_chat_template), and calls generate() with the full sampling knobs - the whole Transformers kitchen sink is exposed. The default system_prompt is a genuinely decent prompt-optimizer: rewrite user input into a complete, expressive prompt, output XML-ish multi-item lists, keep it under 200 words, don't translate text that should render in the image. There's even a line that explicitly tells the model not to censor explicit content - which is the real reason people run local LLMs at all. The default user_prompt ("我要玩原神", "I want to play Genshin") is just the author's sense of humor; ignore it.

Outputs are generated_text (STRING) and history_json_out (STRING). The history output loops back into history_json_in for multi-turn conversation, and images you send in are preserved in that history so a multimodal model can keep "seeing" them across turns.

The inputs that matter

You'll set maybe four things and leave the rest:

• model_instance - from LLMModelLoader. The one non-negotiable.
• user_prompt - what you actually want written/captioned.
• max_new_tokens - default 512; raise it if the model keeps getting cut off mid-prompt.
• temperature (0.7) and top_p (0.95) - the usual randomness pair. For prompt rewriting, if output gets chaotic, drop temperature toward 0.5 before you touch anything else.

The long tail: min_new_tokens, top_k, repetition_penalty, no_repeat_ngram_size, num_beams + length_penalty (only meaningful when beams > 1), do_sample (flip off for deterministic output), and enable_thinking, which passes straight through to apply_chat_template for Qwen3.5-class models that support a thinking toggle.

Installing it

Shared install, worth repeating once: ComfyUI Manager → search comfyui-ez-llm, or:

cd ComfyUI/custom_nodes
git clone https://github.com/spawner1145/comfyui-ez-llm

Restart, then drop a full HuggingFace repo into models/LLM/<name>/. Requirements are just transformers and accelerate - no llama.cpp, no GGUF loader, no external service.

Gotchas worth knowing

It reruns every queue. The should_change toggle (default on) makes the node's IS_CHANGED return NaN, which defeats ComfyUI's caching - that's intentional, because an LLM is non-deterministic and you want a fresh answer each run. But it also means everything downstream reruns, so don't put it in front of a slow sampler and wonder why the graph feels sticky.

The model needs its chat template. LLMModelLoader prints a warning when a model's tokenizer has no chat template and falls back to naive text concatenation. Output quality nosedives; use instruct-tuned models that ship a template.

Size the model to the job. The KB's verdict stands: rewriting a rough idea into a prompt doesn't need a reasoning model, it needs a small obedient one. A 0.6B–3B model on your card, with force_offload on so it unloads between calls, is the sweet spot. And a security note, briefly: this is arbitrary Python in the category that once shipped malware as an "LLM vision" node. It loads only from your local models/LLM folder, but install from known sources and read what you run.