Qwen-Image-Lightning offers distilled versions of the Qwen-Image model, designed to significantly accelerate image generation while retaining complex text rendering capabilities. This project is targeted at researchers and developers seeking faster, yet high-fidelity image synthesis.

How It Works

This project leverages knowledge distillation to create faster, lower-step versions of the Qwen-Image model. By training smaller, faster models to mimic the output of a larger teacher model, it achieves substantial speedups with minimal degradation in quality, particularly for text rendering tasks.

Quick Start & Requirements

• Installation: Install huggingface_hub and download models using huggingface-cli download lightx2v/Qwen-Image-Lightning --local-dir ./Qwen-Image-Lightning.
• Prerequisites: Requires Python environment and base Qwen-Image model setup as per the Qwen-Image project.
• Usage: Evaluation scripts and ComfyUI workflows are provided. Example run commands for 8-step, 4-step, and base models are included.
• ComfyUI: Workflows qwen-image-8steps.json and qwen-image-4steps.json are available, requiring the base model and LoRA weights in the models/loras/ directory.
• Links: Qwen-Image (for base model setup), ComfyUI

Highlighted Details

• Distilled models offer 8-step and 4-step generation options.
• Preserves complex text rendering capabilities.
• Includes evaluation scripts for comparing distilled vs. base models.
• Provides verified ComfyUI workflows for seamless integration.

Maintenance & Community

The project has seen recent releases and updates, indicating active development. Links to community resources are not explicitly provided in the README.

Licensing & Compatibility

Models are licensed under the Apache 2.0 License. This license permits commercial use and linking with closed-source projects, with the caveat that users are responsible for ensuring their generated content complies with applicable laws and ethical guidelines.

Limitations & Caveats

The project acknowledges that neither the distilled nor the base Qwen-Image models consistently generate perfect results, and performance can vary across different prompts and resolutions. A "badcase" example of the distilled model's performance is noted.