Fashion-VDM - A virtual try-on technology using video diffusion models for high-quality dynamic garment visualization.

## Definition of Fashion-VDM Fashion-VDM is an advanced virtual try-on technology developed jointly by Google and the University of Washington. It utilizes video diffusion models (VDM) to generate high-resolution (512px) dynamic try-on videos from a single garment image and a person's video input, maintaining temporal consistency and fine-grained details. ## Developers of Fashion-VDM Fashion-VDM was collaboratively developed by **Google** and the **University of Washington**. ## Technical Methodology Fashion-VDM employs: 1. **Video Diffusion Model (VDM) architecture** for high-quality generation 2. **Split classifier-free guidance** for enhanced conditional control 3. **Progressive temporal training** strategy for efficient long-sequence video generation (64 frames) 4. **3D-Conv and temporal attention blocks** to maintain temporal coherence 5. **Joint image-video training** for data efficiency ## Core Features Key capabilities include: - Generating 64-frame videos at 512px resolution in a single pass - Preserving subject identity and motion fidelity - Supporting multiple conditional inputs (garment-only, person+garment, person+garment+pose) - Demonstrating superior temporal consistency compared to existing methods ## Use Cases Main applications include: 1. **Online retail**: Virtual try-on experiences for e-commerce 2. **Virtual fashion**: Dynamic garment presentation for digital fashion shows 3. **Personalized recommendations**: Customized try-on videos based on user preferences ## Training Methodology The training involves: 1. **Spatial pre-training**: Initial training on image data 2. **Progressive temporal training**: Gradual extension to longer video sequences 3. **Adaptive batch sizing**: Using increasingly longer frame batches for efficient learning ## Access Points Available resources: - [Project Website](https://johannakarras.github.io/Fashion-VDM/) - [Research Paper](https://arxiv.org/abs/2411.00225) - [Supplementary Materials](https://johannakarras.github.io/Fashion-VDM/static/pdf/Fashion_VDM_Supplementary.pdf) - [UBC Benchmark Dataset](https://johannakarras.github.io/Fashion-VDM/static/data/ubc-benchmark.zip) ## Core Features Fashion-VDM outperforms existing methods by: 1. Achieving higher temporal consistency through 3D-Conv and attention mechanisms 2. Generating longer videos (64 frames) at commercial-ready resolution (512px) 3. Maintaining better identity preservation and garment detail fidelity 4. Supporting flexible conditional control via split classifier-free guidance ### Citation sources: - [Fashion-VDM](https://johannakarras.github.io/Fashion-VDM) - Official URL Updated: 2025-04-01