Implementation of VIBE (Visual Instruction Based Editor) as a custom node for ComfyUI. Edit images using natural language instructions (e.g., "make it winter", "change the dog to a cat").

VIBE leverages the efficient Sana1.5-1.6B diffusion model and Qwen3-VL-2B-Instruct for fast, high-quality image manipulation.

✨ Features

Instruction-based Editing: No complex prompting required.
Latent Support: Connect an Empty Latent Image to define output resolution.
Fast Inference: Powered by Sana1.5's linear attention.
Local Model Support: Runs entirely on your hardware.

🚀 Installation

Clone the repository: Navigate to your ComfyUI/custom_nodes folder and run:
```
git clone https://github.com/ato-zen/ComfyUI-VIBE
```
Install dependencies: Open terminal in the ComfyUI-VIBE folder and run:
```
pip install -r requirements.txt
```

📂 Model Setup

This node looks for models in: ComfyUI/models/vibe/

Create the target folder:
```
mkdir -p models/vibe
cd models/vibe
```
Download the weights: Clone from Hugging Face (requires git-lfs):
```
git clone https://huggingface.co/iitolstykh/VIBE-Image-Edit
```

Structure should look like:

ComfyUI/models/vibe/
└── VIBE-Image-Edit/
    ├── model_index.json
    ├── unet/
    ├── vae/
    ├── text_encoder/
    └── tokenizer/

📜 Credits & Acknowledgements

Original Project: VIBE: Visual Instruction Based Editor
Model Authors: Grigorii Alekseenko, Aleksandr Gordeev, Irina Tolstykh, et al.
Based on: Sana and Qwen3-VL.

ComfyUI Node implementation by ato-zen.

License: Apache 2.0