LoRA Fine-tuned Qwen Model for Style Reference Image Edit
Repository: https://huggingface.co/svjack/Qwen_Image_Edit_2509_Sref_Lora
This repository provides a LoRA (Low-Rank Adaptation) fine-tuned variant of the Qwen image model, specialized in generating new images that match the artistic style of reference images while allowing for content transformation. The model combines Qwen3-VL's ability to analyze and describe artistic styles with LoRA's efficient adaptation for style transfer.
✨ Key Features
Style Preservation
- Accurately extracts and replicates artistic characteristics from reference images including color palette, brush strokes, texture, and overall aesthetic style
- Maintains stylistic consistency across different content subjects and compositions
- Supports various art styles including illustration, painting, digital art, and photography
Content Flexibility
- Allows complete content transformation while preserving reference style characteristics
- Enables creative recombination of different subjects within consistent artistic styles
- Supports complex style descriptions and parameter controls for fine-grained adjustments
Prompt Efficiency
- Simple commands trigger precise style transfers and content adaptations
- Supports detailed style weight adjustments for controlling influence strength
🎨 Technical Approach
Style Analysis Pipeline
- Qwen3-VL Analysis: The model first uses Qwen3-VL to analyze reference images and generate detailed style description prompts
- LoRA Adaptation: The fine-tuned LoRA parameters specialize in style transfer and adaptation
- Content-Style Fusion: Combines target content descriptions with extracted style characteristics for final image generation
Method Comparison
This approach differs from traditional style transfer methods by leveraging advanced vision-language understanding for more accurate style characterization and more flexible content-style separation.
🖼️ Transformation Examples
Example 1: Red Parrot
Reference Style Analysis: Red Parrot transformed from a red man with glasses
- target prompt
鹦鹉
- Qwen3 VL Edit prompt
进行下面的修改:
- 将原图中的人物头部及躯干部分替换为一只鹦鹉的形象,保留其轮廓结构并延续原有构图比例;
- 鹦鹉身体覆盖着类似原图服装纹理的黑色底色,并在颈部至背部区域添加红色条纹装饰,颜色走向参照原图衣领设计;
- 保持头盔式装置不变形但调整形态适配鹦鹉头部,新增黄色发光带环绕眼部位置(呼应原图护目镜光效),同时两侧耳罩改为鸟喙侧翼造型;
- 原图蓝色皮肤质感转化为鹦鹉羽毛层次,在面部与颈肩处呈现紫色过渡调并与红橙背景形成高对比度视觉效果;
- 背景仍维持纯正饱和橙红色块面风格,无任何环境元素介入;所有边缘线粗细统一且具平面剪影感,强化矢量插画特征;
- 所有光影处理采用平涂+强描边手法,避免渐变阴影,确保画面呈现出一致性的复古赛博朋克波普艺术表现形式。
Example 1: Cat Playing with Yarn Ball (Stamp Style)
Reference Style Analysis: Vintage stamp design with perforated borders, teal background, and retro texture effects.
- target prompt
猫咪玩毛线球
- Qwen3 VL Edit prompt
进行下面的修改:
- 将原图中的狐狸替换为一只蜷缩着身体、正在玩耍毛线球的小猫;
- 保持原有邮票边框及齿孔样式不变,并延续青绿色背景基调;
- 原红色圆形太阳图案保留并调整大小适配新主体位置,在小猫后方作为视觉焦点;
- 所有文字标签(左上角红圆内汉字、右上角竖排矩形字块等)均沿用相同字体风格与布局结构,但内容需改为符合“猫咪玩毛线球”主题的新字样,如“喵星人”、“绒球乐园”等;
- 背景纹理维持颗粒感质感,确保画面统一复古贴纸效果;
- 毛线球形态从原本无明确对象变为清晰可见且被猫咪互动缠绕的状态;
- 颜色调性适度微调使主色调更偏向暖橘/米白系突出可爱氛围,同时保证橙色日轮仍具辨识度;
- 维持所有阴影过渡与笔触肌理一致,强化图像同质化艺术表达手法。
——最终成品在构图、材质表现力、装饰元素配置等方面完全继承原始画风特征而实现主题转换。
Example 2: Summer Swimming Pool Cube Model
Reference Style Analysis: 3D miniature landscape with precise lighting, material textures, and spatial composition.
- target prompt
夏季游泳池方块模型
- Qwen3 VL Edit prompt
进行下面的修改:
- 将雪地场景完全替换为一个清澈湛蓝的游泳池水面,并保留原图中的立方体地形结构;
- 保持房屋建筑形态不变,但将其屋顶改为泳池边沿风格并移除积雪覆盖层;窗户内暖光仍可维持,作为视觉焦点延续温暖氛围;
- 原有松树全部替换成热带棕榈树,在水中倒影处增加波纹效果模拟水动感;
- 红色卡车被替换为空白白色小船漂浮于泳池中央位置,车轮部分变为船桨样式;
- 替换天空背景由星点夜空转为晴朗蓝天无云状态,同时在上方添加几缕轻盈白云点缀;
- 地面剖开露出下方土质结构时,原有深灰色岩石纹理改绘成浅黄色沙滩质地;
- 所有光影处理统一采用高亮反差表现手法(如反射光泽),确保画面呈现一致性的3D立体质感及材质过渡。
所有元素均依据原始图像的空间布局比例调整尺寸关系,使新构图完整契合同一极简主义微缩景观艺术风格。
Example 3: Golden Retriever in Bubbles
Reference Style Analysis: Dreamy soft-focus photography with bokeh effects, ethereal lighting, and translucent elements.
- target prompt
泡沫中的金毛犬
- Qwen3 VL Edit prompt
进行下面的修改:
- 将原图中的人物主体替换为一只金毛犬,其身体姿态保持类似舞蹈般的舒展弧度,并置于画面中央偏右位置;
- 保留原有的梦幻柔光效果及背景散景(bokeh)结构,在狗狗周围添加大量半透明气泡状元素,使其仿佛悬浮于水中或雾霭之中;
- 调整服装部分——原本粉色纱质舞裙被替换成具有光泽质感的浅金色绒面材质,贴合狗狗体型并覆盖全身;袖口设计改为短而蓬松的流苏样式,延续轻盈飘逸感;
- 颜色基调从粉蓝渐变过渡至暖橙调为主导,但依然维持柔和朦胧氛围,强化光影在皮毛上的反射高光区域;
- 原有面部轮廓消失后新增拟人化侧脸线条暗示方向性动作,同时增强眼部发光特效模拟光源穿透水体时形成的粒子折射现象;
- 整个构图仍采用低角度仰拍视角加强视觉纵深,边缘模糊处理依旧存在且保持一致虚实对比层次。
所有改动均沿袭原始图像所呈现的艺术风格特征,包括光线运用方式、纹理表现手法以及情绪渲染手段等关键要素未作偏离。
Example 4: Yellow-Green Firefighter
Reference Style Analysis: Surreal portrait style with blurred edges, chromatic transitions, and fluorescent color effects.
- target prompt
黄绿色的消防员
- Qwen3 VL Edit prompt
进行下面的修改:
- 将原图中的角色主体替换为一个穿着黄色制服并配有红色条纹及肩章的消防员形象,其服装结构包括胸前徽章式图案(类似星形但风格化)、腰带以及臂部护甲设计;
- 对该消防员的形象在视觉呈现上保持原有模糊边缘处理,并延续画面中心聚焦于人物的姿态特征;
- 背景保留原有的梦幻般光晕效果,在此基础上增强彩虹色谱过渡——从左上方蓝色渐变至右下方橙红紫混合区域;同时强化前景地面处流动般的彩色折射纹理,使其具有透明薄膜质感;
- 通过调整饱和度提升部分高亮区如胸口徽章与袖口反光部位的颜色浓度,使整个图像呈现出更强烈的荧光感;
- 在所有颜色层面上统一应用偏冷色调主调基础上叠加暖色点缀的方式,确保新加入的人物及其服饰元素仍符合原始场景所具有的超现实光影氛围。
以上操作完全沿袭了源图的艺术表现手法,未引入任何非同质化的特效或构图方式。
Example 5: Arab Knight in Constantinople (Comic Style)
Reference Style Analysis: Bold comic art style with sharp edges, flat color areas, and strong contrast.
- target prompt
漫画阿拉伯骑士在君士坦丁堡
- Qwen3 VL Edit prompt
进行下面的修改:
- 将原图中的修女服饰替换为带有盔甲元素的阿拉伯风格骑士装束,包括肩部护甲、胸甲及腿部铠片,并保留原有紫色调作为主色系;
- 原图中佩戴的眼镜保持不变但调整位置使其更符合骑士形象;头巾改为装饰性头饰并增加金色纹样细节;
- 手势从比耶手势变为握剑姿态(右手持剑),左手自然下垂且仍带手套,剑柄部分采用金蓝色镶嵌设计并与角色服装颜色呼应;
- 背景由教堂尖塔结构改绘成具有伊斯兰建筑特征的城市轮廓——如拱形门廊、圆顶式屋顶以及多角塔楼形态,在左侧背景加入喷溅状红蓝颜料效果增强视觉张力;
- 画面右侧新增一尊类似清真寺宣礼塔的高耸建筑物剪影,其顶部呈现新月符号造型;
- 整体线条依旧延续锐利边缘+平涂块面构成的动漫插画质感,仅通过局部构图重组实现主题转换而不破坏原有的艺术语言一致性。
所有改动均围绕同一套鲜明对比强烈的冷暖撞色体系展开,确保最终图像呈现出统一而富有冲击感的新场景设定。
⚙️ Usage Instructions
- refer on workflow in
- https://huggingface.co/svjack/Qwen_Image_Edit_2509_Sref_Lora/resolve/main/qwen_image_edit_2509_vl_sref.json
🔧 Technical Details
- Base Model: Qwen Image Edit 2509 Model
- LoRA Rank: 32
- Training Data: 2,000+ style-content pairs across diverse artistic genres
- Style Analysis: Qwen3-VL for automated style characterization and prompt generation
- Specialized Capabilities:
- Accurate style element extraction and replication
- Content-style disentanglement and recombination
- Multi-style blending and interpolation
- Style strength modulation
Optimized for: Single-GPU inference (NVIDIA 20GB+ recommended) with batch processing support for style-consistent image series generation.
🎯 Best Practices
Reference Image Quality
- Use high-quality, stylistically consistent reference images
- Ensure reference images have clear stylistic characteristics
- Avoid references with conflicting or ambiguous styles
Prompt Formulation
- Clearly separate style and content instructions
- Use the model's specialized style vocabulary
- Specify style strength parameters based on desired outcome
This model is particularly valuable for creators, designers, and artists looking to maintain stylistic consistency across different subjects or explore creative variations within established visual styles.
Note: This model is designed for creative style transfer applications. All copyrights for reference images and generated content belong to their respective owners.











