PaperlayerView exact revision on arXiv

Source-linked AI summary

Eevee: Towards Close-up High-resolution Video-based Virtual Try-on

Jianhao Zeng, Yancheng Bai, Ruidong Chen, Xuanpu Zhang, Lei Sun, Dongyang Jin, Ryan Xu, Nannan Zhang, Dan Song, Xiangxiang Chu

arXiv:2511.18957v2cs.CV

TL;DR

Existing video virtual try-on datasets often lack close-up videos, high resolution, and detailed garment inputs, limiting the capture of garment details. Eevee provides high-resolution garment and video data, and VGID evaluates garment consistency. The benchmark identifies texture and structural preservation disparities among current methods.

  • Problem

    Existing datasets commonly lack close-up video, high resolution, and detailed garment images, while current metrics do not adequately assess garment consistency.

    Source: §1 · p. 2, §3.1 · p. 3

  • Method

    Eevee provides high-resolution full-shot and close-up videos with detailed garment images and descriptions, alongside VGID, which compares garment-region features with input-image features.

    Source: §1 · p. 2

  • Results

    The benchmark of recent models reveals significant disparities in texture and structural preservation.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The Eevee dataset and VGID benchmark establish a baseline for future research in high-fidelity virtual try-on.

    Source: §1 · p. 2

    VGID is reported exclusively for dresses because its soft masks can include adjacent garments in lower- and upper-body categories.

    Source: §3.3 · p. 5

Abstract

from arXiv · show

Video virtual try-on technology provides a cost-effective solution for creating marketing videos in fashion e-commerce. However, its practical adoption is hindered by two critical limitations. First, the reliance on a single garment image as input in current virtual try-on datasets limits the accurate capture of realistic texture details. Second, most existing methods focus solely on generating full-shot virtual try-on videos, neglecting the business's demand for videos that also provide detailed close-ups. To address these challenges, we introduce a high-resolution dataset for video-based virtual try-on. This dataset offers two key features. First, it provides more detailed information on the garments, which includes high-fidelity images with detailed close-ups and textual descriptions; Second, it uniquely includes full-shot and close-up try-on videos of real human models. Furthermore, accurately assessing consistency becomes significantly more critical for the close-up videos, which demand high-fidelity preservation of garment details. To facilitate such fine-grained evaluation, we propose a new garment consistency metric VGID (Video Garment Inception Distance) that quantifies the preservation of both texture and structure. Our experiments validate these contributions. We demonstrate that by utilizing the detailed images from our dataset, existing video generation models can extract and incorporate texture features, significantly enhancing the realism and detail fidelity of virtual try-on results. Furthermore, we conduct a comprehensive benchmark of recent models. The benchmark effectively identifies the texture and structural preservation problems among current methods.

1. Introduction

Existing video virtual try-on datasets and methods offer limited resolution, garment detail, and close-up video, constraining practical use. Eevee adds high-resolution garment and video data, VGID for garment consistency, and a benchmark of recent models.

1. Introduction

  • Existing datasets commonly provide only a full-shot video, low resolution, and a single flat garment image, limiting close-up detail and texture capture.Vendors can supply multi-angle garment detail shots, but this information remains largely underused.Source: §1 · p. 2
  • Eevee provides full-shot and close-up videos, detailed garment images and descriptions, and high resolution for virtual try-on tasks.The comparison with ViViD reports clearer details and superior texture fidelity for Eevee.Source: §1 · p. 2, Figure 2 · p. 2
  • VGID measures garment consistency by comparing deep features from generated garment regions with features from the input garment image.It is proposed because VFID and VBench primarily assess overall video quality and realism, not garment consistency.Source: §1 · p. 2
  • Current methods show significant disparities in preserving garment texture and structure in a benchmark using Eevee and VGID.The evaluation also uses VFID, SSIM, LPIPS, and VBench.Source: §1 · p. 2

2. Related Work

Video-based virtual try-on research has begun exploring architectures for synthesizing garments on moving people, but dataset limitations constrain detail and resolution. Earlier image datasets expanded resolution and scale, while video datasets introduced temporal data but still lack close-up views and detailed garment inputs.

2.1. Video-based Virtual Try-on

  • Video-based try-on methods explore U-Net and DiT architectures for spatial and temporal consistency, but typically generate low-resolution, full-shot videos lacking garment detail.Examples include RealVVT [32], CatV2TON [10], and MagicTryon.Source: §2.1 · p. 3

2.2. Virtual Try-on Datasets

  • Early virtual try-on datasets centered on static, image-based collections: VITON [20] provided paired images of a person and target in-shop garment, and VITON-HD [8] increased resolution to 1024 × 768.Source: §2.2 · p. 3
  • Video datasets VVT and ViViD [16] add temporal data, but VVT has low resolution and small scale, while ViViD lacks close-up video and detailed garment images.ViViD uses one flatlay image per garment, limiting capture of textures, material properties, and deformation details.Source: §2.2 · p. 3

3. Eevee

Eevee adds close-up try-on videos and detailed garment data to a high-resolution dataset, while its VACE fine-tuning pipeline incorporates garment-detail images. The paper also proposes VGID for garment consistency evaluation, reporting the metric only for dresses because masks can include adjacent garments in other categories.

3.1. Dataset

  • Eevee pairs full-shot and close-up try-on videos with detailed garment images and descriptions, addressing dataset gaps in viewing angles, garment detail, and resolution.It contains 9,364 pairs, with images up to 2400 × 1800 and videos up to 1088 × 816.Source: §3.1 · pp. 3–4
  • Eevee is the first dataset described here to provide both detailed garment images and close-up try-on videos, with the highest image and video resolutions available for the task.Source: §3.1 · p. 4

3.2. Method

  • VACE fine-tuning encodes garment and detail images as reference latents and prepends them to the video latents used to form Context Tokens.The pipeline also encodes reactive and inactive video frames separately and incorporates a latent representation of the mask.Source: §3.2 · pp. 4–5

3.3. Metric

  • VGID is reported only for dresses because DINO-V2 soft masks can include adjacent garments in upper- and lower-body categories.Source: §3.3 · p. 5
  • VGID compares masked, spatially averaged deep features from a garment image and generated frame using cosine similarity, then averages scores across video frames.The feature masking is intended to suppress background features, and the metric measures semantic garment fidelity such as whether a plaid pattern remains recognizably plaid.Source: §3.3 · p. 5

4. Experiments

Experiments evaluate models on Eevee across full-shot and close-up videos, paired and unpaired settings, and two resolutions. Fine-tuned VACE performs strongly on several metrics and qualitative comparisons, while ablations show that detailed garment images are most effective alongside LoRA fine-tuning.

4.1. Implementation Details

  • Experiments use 49 input and output frames, paired and unpaired configurations, and resolutions of 1088 × 816 and 832 × 624; metrics cover image quality, video quality, and garment consistency.Results report weighted averages across dresses, upper-body, and lower-body categories.Source: §4.1 · pp. 5–6, §4.2 · p. 6

4.2. Quantitative Comparison

  • MagicTryon leads on VFIDI and LPIPS, while fine-tuned VACE leads on the other metrics, particularly in the paired setting.The experiments also report performance decreases for all models on the more difficult high-resolution and close-up generation tasks.Source: §4.2 · pp. 6–7, Table 2 · p. 6

4.3. Qualitative Comparison

  • Fine-tuned VACE reproduces input garment patterns and looks more realistic in close-up results, while ViViD and MagicTryon often fail to preserve patterns and original VACE tends to paste garments onto the target area.VACE can also produce solid white clothing or omit human body parts, which contributes to poor quantitative metrics.Source: §4.3 · p. 7, Figure 5 · p. 7

4.4. Ablation Study

  • Adding detailed garment images alone has negligible impact, but the combination of LoRA fine-tuning and detailed images outperforms LoRA fine-tuning without them.The authors conclude that adaptation is needed for the model to exploit the images’ high-fidelity information.Source: §4.4 · p. 7, Table 3 · p. 8
  • In the ablation visuals, only fine-tuning enables VACE to generate a human, and combining it with detailed images produces a more natural result with clearer, more faithful textures.The original model generated videos containing only the garment; adding detailed images without fine-tuning made output more natural but still omitted the person.Source: §4.4 · pp. 7–8, Figure 6 · p. 8
  • In a 53-person study, raters preferred detailed guidance over the baseline for texture (79.25%), pattern (83.02%), and overall quality (84.91%).The authors interpret this subjective evaluation as confirming the detailed image’s importance for preserving garment identity.Source: §4.4 · p. 8

5. Conclusion

The paper introduces Eevee, a high-resolution virtual try-on dataset with full-shot and close-up videos and detailed garment images. It also proposes VGID for garment consistency and benchmarks recent models to establish a baseline for future research.

5. Conclusion

  • Eevee provides high-resolution full-shot and close-up videos with corresponding detailed close-up garment images, and VGID measures garment consistency in a benchmark of recent models.The paper presents the dataset and metric as addressing low resolution, limited video views, and difficulty capturing fine-grained garment textures from single input images.Source: §5 · p. 8

Related papers

Loading 2511.18957v2…