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

1. Clone this repository into your ComfyUI custom_nodes folder:

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

2. Install dependencies:

   cd ComfyUI_LFM2-350M
   pip install -r requirements.txt
   

3. 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