IMAGGarment-1: Fine-Grained Garment Generation for Controllable Fashion Design

IMAGGarment-1, a new framework from researchers at Nanjing University of Science and Technology and Nanjing University, is making waves in the world of digital fashion design. This innovative system enables high-fidelity garment synthesis with precise control over silhouette, color, and logo placement, addressing the limitations of existing methods that often struggle with multi-conditional controllability.

What Changed?

IMAGGarment-1 introduces a two-stage training strategy to tackle the challenges of personalized fashion design. Unlike previous approaches that can only handle single-condition inputs (like just a sketch or a color palette), IMAGGarment-1 excels in generating garments based on multiple conditions simultaneously. This means designers can specify detailed requirements, such as specific silhouettes, colors, and logo placements, all within a unified framework.

Key Technical Details

• Two-Stage Training Strategy:

  • Global Appearance Model (Stage I): This model encodes the overall silhouette and color using a mixed attention module and a color adapter. It generates a coarse latent representation of the garment.
  • Local Enhancement Model (Stage II): This stage refines the global appearance by integrating user-defined logos and spatial constraints, ensuring accurate placement and visual consistency.

• Mixed Attention Module: Combines both local and global attention mechanisms to capture fine-grained details while maintaining overall coherence.

• Color Adapter: Adjusts the color palette to match the specified conditions, ensuring high color fidelity in the final output.

• Adaptive Appearance-Aware Module: Injects user-defined logos and spatial constraints into the latent representation, enabling precise control over logo placement.

Dataset: GarmentBench

To train and evaluate IMAGGarment-1, the researchers released a large-scale dataset called GarmentBench. This dataset includes over 180K garment samples paired with multi-level design conditions, such as sketches, color references, logo placements, and textual prompts. The diversity of this dataset ensures that the model can handle a wide range of input conditions and generate high-quality outputs.

Performance Benchmarks

Extensive experiments demonstrate that IMAGGarment-1 outperforms existing baselines in several key areas:

• Structural Stability: The generated garments maintain consistent and realistic structures, even with complex designs.
• Color Fidelity: The color adapter ensures that the final images match the specified color palettes accurately.
• Local Controllability: User-defined logos and spatial constraints are accurately placed, enhancing the overall design quality.

Example Applications

IMAGGarment-1 has a wide range of applications in both fashion design and digital apparel:

• Fashion Designers: Can quickly generate multiple design iterations based on client specifications.
• E-commerce Platforms: Can create realistic product images without the need for physical prototypes.
• Virtual Try-On Systems: Can synthesize garments with user-specific measurements and preferences.

Implementation Notes

The code and model for IMAGGarment-1 are available on GitHub, making it accessible to researchers and practitioners. The framework is built using PyTorch, and detailed documentation is provided to help users get started.

Conclusion

IMAGGarment-1 represents a significant step forward in fine-grained garment generation. By addressing the challenges of multi-conditional controllability, this framework opens up new possibilities for personalized fashion design and digital apparel applications. Whether you're a fashion designer looking to streamline your workflow or a researcher exploring the frontiers of AI-generated content, IMAGGarment-1 is worth checking out.