Qwen-Image-i2L (Image to LoRA)

Revolutionary Image to LoRA model that takes images as input and outputs trained LoRA models, enabling instant style transfer and content preservation

Image to LoRA

Style Transfer

LoRA Generation

DiffSynth Studio

Multi-Model System

Visit Resource

Overview

Qwen-Image-i2L is an innovative model that takes images as input and directly outputs LoRA weights trained on those images. The system includes four specialized models: i2L-Style for style transfer, i2L-Coarse and i2L-Fine for content preservation, and i2L-Bias for output alignment with Qwen-Image aesthetics.

Features

Instant LoRA generation from input images without traditional training
Qwen-Image-i2L-Style: 2.4B parameters for effective style extraction and transfer
Qwen-Image-i2L-Coarse: 7.9B parameters for content preservation with SigLIP2, DINOv3, Qwen-VL encoders
Qwen-Image-i2L-Fine: 7.6B parameters with 1024x1024 resolution for detail capture
Qwen-Image-i2L-Bias: 30M static LoRA for Qwen-Image style alignment
Combined Coarse+Fine+Bias mode for high-fidelity content and detail preservation
Supports style transfer with minimal input images
Can serve as initialization weights for accelerated LoRA training

Images

Input image for abstract vector style transfer

Generated image using i2L-Style model with abstract vector style

Input image for black and white sketch style transfer

Generated image using i2L-Style model with sketch style

Installation

Download models from HuggingFace and use with DiffSynth-Studio framework or compatible diffusion pipelines

Usage

For style transfer, use i2L-Style with a few unified style images. For content preservation, combine i2L-Coarse, i2L-Fine, and i2L-Bias models. All showcase examples use random seed 0.

Requirements

Python 3.8+
PyTorch with CUDA support
DiffSynth-Studio or compatible framework
8GB+ VRAM for Style model, 16GB+ for Coarse+Fine+Bias combination