Qwen-Image-Edit OmniConsistency LoRA: Artistic Style Transfer Model

This repository provides a LoRA (Low-Rank Adaptation) fine-tuned variant of the Qwen-Image-Edit model integrated with OmniConsistency technology, optimized for transforming images into 22 distinct artistic styles while preserving content consistency and facial details.

🎨 Supported Art Styles

Style Category	Example Prompt	Visual Characteristics
3D Chibi Style	`transform it into 3D Chibi style`	Exaggerated cute proportions with three-dimensional rendering and soft shading
American Cartoon Style	`transform it into American Cartoon style`	Bold outlines, vibrant colors, and exaggerated expressions typical of Western animation
Chinese Ink Style	`transform it into Chinese Ink style`	Flowing brushstrokes, monochromatic tones, and traditional shan shui aesthetics
Clay Toy Style	`transform it into Clay Toy style`	Matte textures with visible fingerprints and soft plasticine-like appearance
Fabric Style	`transform it into Fabric style`	Woven textile appearance with stitch details and cloth-like folds
Ghibli Style	`transform it into Ghibli style`	Soft watercolor-like backgrounds, expressive eyes, and whimsical Studio Ghibli aesthetic
Irasutoya Style	`transform it into Irasutoya style`	Clean vector graphics with flat colors and simple shapes (Japanese clipart style)
Jojo Style	`transform it into Jojo style`	Dynamic "bizarre" poses, exaggerated muscles, and dramatic manga shading
LEGO Style	`transform it into LEGO style`	Blocky construction with cylindrical hands and studded surfaces
Line Style	`transform it into Line style`	Minimalist continuous-line drawings with negative space emphasis
Macaron Style	`transform it into Macaron style`	Pastel colors with soft gradients and candy-like textures
Oil Painting Style	`transform it into Oil Painting style`	Visible impasto brushstrokes and rich pigment textures
Origami Style	`transform it into Origami style`	Geometric folded paper appearance with crisp edges
Paper Cutting Style	`transform it into Paper Cutting style`	Silhouette art with intricate negative space patterns
Picasso Style	`transform it into Picasso style`	Cubist fragmentation and abstract facial rearrangements
Pixel Style	`transform it into Pixel style`	8-bit/16-bit retro game aesthetic with visible square pixels
Poly Style	`transform it into Poly style`	Low-polygon 3D models with flat-shaded triangular facets
Pop Art Style	`transform it into Pop Art style`	Ben-Day dots, bold colors, and high-contrast comic book styling
Rick Morty Style	`transform it into Rick Morty style`	Squiggly lines, grotesque proportions, and adult swim animation style
Snoopy Style	`transform it into Snoopy style`	Simple black-and-white comic strip aesthetic with round features
Vector Style	`transform it into Vector style`	Clean geometric shapes with gradient fills and sharp edges
Van Gogh Style	`transform it into Van Gogh style`	Swirling brushwork, thick impasto, and post-impressionist color fields

🖼️ Style Transformation Examples

1. 3D Chibi Style (包拯)

Source Image	Target Image without LoRA	Target Image with OmniConsistency LoRA

Prompt: `transform it into 3D Chibi style`	Issues: Weak rendering	Advantages: relatively exaggerated 3D rendering

2. Jojo Style (叶卡捷琳娜二世)

Source Image	Target Image without LoRA	Target Image with OmniConsistency LoRA

Prompt: `transform it into Jojo style`	Issues: Weak pose dynamics, inconsistent shading	Advantages: Enhanced "bizarre" poses, dramatic manga lighting

3. Snoopy Style (土耳其神枪手)

Source Image	Target Image without LoRA	Target Image with OmniConsistency LoRA

Prompt: `transform it into Snoopy style`	Issues: Missing signature details, inconsistent linework	Advantages: Preserved weapon details, authentic comic strip aesthetic

4. Pop Art Style (猫咪太师大壮和西洋蔡)

Source Image	Target Image without LoRA	Target Image with OmniConsistency LoRA

Prompt: `transform it into Pop Art style`	Issues: Weak Ben-Day dots, color bleeding	Advantages: Crisp dot patterns, vibrant color separation

✨ Technical Workflow

graph LR
    A[Original Image] --> B(Qwen-Image Encoder)
    B --> C{OmniConsistency Module}
    C -->|Style Prompt| D[LoRA Adapter Bank]
    D -->|3D Chibi| E["<img src='https://huggingface.co/datasets/svjack/Xiang_idol_Kontext_OmniConsistency_lora_Images/resolve/main/chibi_example.jpg' width='120'/>"]
    D -->|Jojo| F["<img src='https://huggingface.co/datasets/svjack/Premier_Zhou_OmniConsistency_Images/resolve/main/jojo_example.jpg' width='120'/>"]
    D -->|Snoopy| G["<img src='https://example.com/snoopy_example.jpg' width='120'/>"]
    D -->|Pop Art| H["<img src='https://example.com/popart_example.jpg' width='120'/>"]

Key Features:

Consistency Preservation: Maintains facial features and complex scene details
Flexible Layout Control: Supports creative structural changes like chibi proportions
Multi-Style Compatibility: Plug-and-play integration with any style LoRA module

Explore More Examples:

https://huggingface.co/datasets/svjack/Xiang_hoodies_Qwen_Image_Edit_OmniConsistency_lora_Images

Model Details

Base Architecture: Qwen-Image (ViT-H/16 visual encoder + Qwen-7B text encoder)
LoRA Configuration:
- Rank: 32 (visual), 32 (text)
Training Data: 2,600 curated image pairs across 22 styles
Model File: https://huggingface.co/svjack/Qwen_Image_Edit_OmniConsistency_lora

Acknowledgement: Incorporates OmniConsistency research from National University of Singapore