Skip to main content
Back

Alibaba's Eevee Dataset Fixes Blurry AI Fashion Try-On at 2400x1800

A new Alibaba dataset pushes video try-on to 2400x1800 images with close-up shots, plus a DINO-V2 metric that scores texture fidelity.

Pro

  • Alibaba's Amap team releases Eevee, a 9,364-pair video try-on dataset at up to 2400x1800 image resolution.
  • First dataset to include both full-shot and close-up try-on videos plus multi-angle detailed garment shots.
  • New VGID metric uses DINO-V2 features and emergent segmentation to score garment texture fidelity in generated frames.
  • Fine-tuning VACE with LoRA plus detailed garment images beats MagicTryOn on SSIM and VGID.
  • Ablation shows detailed images only help when combined with fine-tuning; raw injection into base VACE has negligible effect.
  • Code, weights, and dataset available at github.com/AMAP-ML/Eevee and Hugging Face.

Fashion e-commerce runs on video, but the models that generate try-on clips have been stuck training on blurry, full-body footage where you can barely tell wool from polyester. A team from Alibaba's Amap group, Tianjin University, and the Shenzhen Institutes of Advanced Technology is trying to fix that with Eevee, a high-resolution dataset that adds close-up videos and multi-angle garment shots, along with a new metric built to catch when a generated jacket loses its plaid.

The blur problem in virtual try-on

Video-based virtual try-on takes a garment image plus a source video of a person and synthesizes a new clip where that person wears the target garment while their movement and identity stay intact. Existing public datasets typically contain footage from a single global perspective, while practical applications often require close-ups on specific areas. They also represent each garment with a single flat in-shop image, overlooking a readily available asset: merchants can provide detailed close-up shots. Most of these datasets suffer from low resolution too, which caps how sharp any generated result can ever be.

The scale of the resolution gap is stark. The widely used ViViD dataset tops out at 1380x920 for images and 832x624 for videos, while VVT sits at 256x192. At those sizes, fabric weave, stitching, and small prints disappear into JPEG mush before a model ever sees them.

What Eevee actually contains

Eevee ships 9,364 high-resolution video-based try-on pairs, with images up to 2400x1800 and videos up to 1088x816, making it the highest-resolution collection currently available for this task. The dataset splits across three garment types and totals 28,092 images and 1.9 million video frames: