Extensions/ComfyUI-Prompt-Rewriter
ComfyUI Extension

ComfyUI-Prompt-Rewriter

Prompt Rewriter for ComfyUI using llama.cpp for local inference. Supports LLMs, VLMs, and both Instruct and Thinking models.

By BigStationW·Created 9 months ago·Updated 5 months ago· 20
BigStationW/ComfyUI-Prompt-Rewriter
Nodes2
On cloudLocal install
CategoryPrompt Manager
Stars20
Updated5 months ago

Nodes (2)

Prompt Rewriter Options
Prompt Manager
Prompt Rewriter
Prompt Manager
Readme

ComfyUI Prompt Rewriter

This is a custom fork of this repository incorporating some enhancements.

Use case

This node can rewrite your prompts with the help of a chosen instruct (or thinking) LLM/VLM (GGUF).

<img width="700" alt="AA" src="https://github.com/user-attachments/assets/c2b4de17-0121-4f9f-a845-3c0c52f74bce" />

Installation

  1. Navigate to the ComfyUI/custom_nodes folder, open cmd and run:
git clone https://github.com/BigStationW/ComfyUI-Prompt-Rewriter
  1. Navigate to the ComfyUI\custom_nodes\ComfyUI-Prompt-Rewriter folder, open cmd and run:
..\..\..\python_embeded\python.exe -s -m pip install -r "requirements.txt"

Backend: Vulkan -> Works on all GPUs

If you have Windows, open cmd and run:

winget install llama.cpp

If you have another OS, you can refer to this.

To update llama.cpp, open cmd and run:

winget upgrade llama.cpp

Backend: CUDA -> Specialized for Nvdia

Double click the Install llama.cpp (CUDA).bat file.

To update your version, you can run that file again.

Instruct/Thinking LLMs

  1. Navigate to the ComfyUI\models folder and create a folder named "LLM"
  2. Navigate to the ComfyUI\models\LLM folder and create a folder named "gguf"
  3. Navigate to the ComfyUI\models\LLM\gguf folder and place your chosen GGUF LLM file there.

For example you can go for this (Instruct model):

  • https://huggingface.co/Qwen/Qwen3-4B-GGUF

Or something like this (thinking model):

  • https://huggingface.co/unsloth/Qwen3-4B-Thinking-2507-GGUF

You can even go for uncensored LLMs, like that one for example:

  • https://huggingface.co/mradermacher/Josiefied-Qwen3-8B-abliterated-v1-GGUF

Usage

It should look like this.

<img width="700" alt="image" src="https://github.com/user-attachments/assets/e29ec62a-8770-41cf-99ed-7438a6fd4894" />

An example workflow (for Z-image turbo) can be found here.

PS: The Display Any (rgthree) node can be found here.

MultiGPU and offloading

This node allows you to split the model into your GPUs and the CPU with the gpu_layers placeholder.

For example:

  • Empty placeholder -> All layers go to the first GPU (default)
  • gpu0:0.7 -> 70% to GPU:0, 30% to CPU
  • gpu0:0.5, gpu1:0.4 -> 50% GPU:0, 40% GPU:1, 10% CPU
<img width="700" alt="image" src="https://github.com/user-attachments/assets/54bd5a5c-73c5-4271-856b-997bcb0a48c5" />

Image inputs

For Vision Language Models (VLMs), you can add up to 5 images to the Prompt Generator Options node.

  1. Download a VLM gguf file and put it to the ComfyUI\models\LLM\gguf folder, like that one for example:
  • https://huggingface.co/unsloth/Qwen3-VL-4B-Thinking-GGUF
  1. Download its mmproj file and put it to the ComfyUI\models\LLM\gguf folder.
  • https://huggingface.co/unsloth/Qwen3-VL-4B-Thinking-GGUF/blob/main/mmproj-BF16.gguf
  1. You have to rename "mmproj-BF16.gguf" to "Qwen3-VL-4B-Thinking-mmproj-BF16.gguf"

At the end it'll look like this:

<img width="700" alt="image" src="https://github.com/user-attachments/assets/a2f602a4-f156-49b8-b70b-4f94b7900d95" />