ComfyUI_LFM2-350M

ComfyUI_LFM2-350M
★ 2

模型加载prompt增强长序列模型生成参数
为 ComfyUI 加载并使用 LFM2-350M 模型，作为 z-image turbo 等长序列模型的 prompt 增强器，支持 HuggingFace/本地加载、torch.compile 权重修复及生成参数配置。
💡 为 z-image turbo 等长 token 模型生成或优化提示词
🍴 2 Forks💻 Python🔄 2026-01-11
📦
网盘下载
复制链接后前往夸克网盘下载
https://pan.quark.cn/s/79aaff81621b
📦 requirements.txt
transformers
torch
accelerate
safetensors
📄 README

ComfyUI LFM2-350M Node

A custom node for ComfyUI to load and use my custom tuned LFM2-350M model trained on 60k base pairs to work as a prompt enhancer for z-image turbo or any other long token model.

Features

  • Load from HuggingFace or a local path
  • Supports fine-tuned models saved with torch.compile() (automatically fixes _orig_mod. weight prefix)
  • Configurable generation parameters (temperature, top_p, top_k, min_p, repetition_penalty)
  • System prompt and user prompt inputs

    • Installation

  • Clone this repository into your ComfyUI custom_nodes folder:

    • “`bash

    cd ComfyUI/custom_nodes

    git clone https://github.com/marduk191/ComfyUI_LFM2-350M.git

    “`

  • Install dependencies:

    • “`bash

    cd ComfyUI_LFM2-350M

    pip install -r requirements.txt

    “`

  • Restart ComfyUI.

    • Nodes

    LiquidAI LFM-2-350M Loader

    Loads the model and tokenizer.

    | Input | Description |

    |——-|————-|

    | repo_id | HuggingFace repository ID (default: marduk191/lfm2-350m-dp-marduk191) |

    | local_path | Optional local path to a pre-downloaded/fine-tuned model |

    | precision | Model precision: bf16, fp16, fp32, or auto |

    | device | cuda or cpu |

    LiquidAI LFM-2-350M Generator

    Generates text based on prompts.

    | Input | Description |

    |——-|————-|

    | model_context | Connect to Loader’s model output |

    | tokenizer | Connect to Loader’s tokenizer output |

    | system_prompt | System instructions for the model |

    | prompt | User input text |

    | max_new_tokens | Maximum tokens to generate (1-4096) |

    | temperature | Randomness control (default: 0.3) |

    | top_p | Nucleus sampling parameter |

    | top_k | K sampling parameter |

    | min_p | Minimum probability (default: 0.15) |

    | repetition_penalty | Penalty for repeating tokens (default: 1.05) |

    Recommended Parameters

    For best results with LFM2-350M:

  • temperature: 0.3
  • min_p: 0.15
  • repetition_penalty: 1.05

    • Requirements

  • Python 3.10+
  • PyTorch 2.0+
  • Transformers 4.55+
  • CUDA GPU recommended

    • Credits

  • Model: LiquidAI – LFM2-350M on HuggingFace
  • ComfyUI: comfyanonymous

    • License

    MIT License