An experimental LoRA fine-tuned model based on Qwen-Image-Edit-2511, designed for style transfer and artistic transformation. Trained with over 400 image pairs for 22,000 steps at rank 32, using a learning rate of 0.0001 and training resolution of 1024, achieving approximately 15 seconds per iteration on an NVIDIA RTX 4090.
🔍 工作原理
给定:
一张源图像(例如,一个人的照片)
一张参考图像(例如,卡通、线稿、插画等)
微信图片_20251230170631_361_150
模型将参考图像的视觉风格应用于源图像,同时保留其结构和构图。虽然它仍处于早期阶段,但在ComfyUI中已经能够对部分风格产生良好的效果。
✅ 示例:将cosplay照片转化为Lacoste鳄鱼素描风格、色彩斑斓的波普艺术狗或极简主义冬季卡通——只需一键。
🔍 How It Works
Given:
A source image (e.g., a photo of a person)
A reference image (e.g., cartoon, line art, illustration, etc.)
The model applies the visual style of the reference image to the source image while preserving its structure and composition. Although still in its early stages, it has shown promising results for certain styles in ComfyUI.
✅ Example: Turn a cosplay photo into a Lacoste-style crocodile sketch, a colorful pop-art dog, or a minimalist winter cartoon — all with one click.
🛠️ 使用指南
在 ComfyUI 中:
加载你的源图像(例如,一张照片)。
加载一个风格参考图像。
应用此 LoRA,强度设置为0.6–1.0。
使用图像到图像或修复节点生成结果。
尽管该模型在某些风格上表现良好,但它仍在开发中,可能会遇到一些局限性。
In SD WebUI:
Load the LoRA via "Load LoRA" tab.
Set LoRA weight to 0.7–1.0.
Use with img2img mode and a reference image as input.
💡 Tip: For best results, use references with similar aspect ratios and compositions.
🧪 训练详情
基础模型: Qwen-Image-Edit-2511
训练方法: LoRA 微调(Rank=32, Alpha=16)
数据集: 自定义精选的艺术风格数据集(线稿、卡通、波普艺术、超现实主义等),包含超过400组图像
Epochs: 22000 步
Batch Size: 1
学习率: 1e-4
优化器: AdamW
训练硬件: NVIDIA RTX 4090,大约每迭代15秒
⚠️ 局限性与未来工作