EquiFashion-DB (Mini)

EquiFashion-DB (Mini) is a compact subset of EquiFashion-DB with aligned multimodal signals for controllable fashion generation: image, text, pose, sketch, and fabric.

Full Dataset Public at: https://drive.google.com/file/d/13TS1U0IY8oG1gjMvsGCQCXrxLxm2SH-Z/view?usp=drive_link

Structure (current EquiFashion_DB/)

EquiFashion_DB/
├── train/               # training images
├── test/                # test images
├── train_pose/          # pose assets (json/ + pose/ visualizations)
├── train_sketch/        # extracted Canny sketch maps (PNG)
├── train_fabric/        # extracted fabric texture patches (PNG)
├── train.json           # train captions (list of {gt, caption})
├── test.json            # test captions (list of {gt, caption})
└── train_pose.json      # train captions + pose path (list of {gt, caption, pose})

Annotation format (as provided)

Train captions (train.json)

{ "gt": "009292_0.jpg", "caption": "Sweater, Commute, Homewear, ..." }

Train captions + pose path (train_pose.json)

{ "gt": "009292_0.jpg", "caption": "Sweater, ...", "pose": "train_pose/pose/009292_0.jpg" }

Pose keypoints JSON (train_pose/json/<gt_stem>.json)

• Key candidate: list of ([x, y, confidence, joint_index])

{
  "candidate": [[282.0, 3.0, 0.54, 0.0], [247.0, 58.0, 0.92, 2.0]]
}

Modalities

• Image: train/<gt> and test/<gt>
• Text: train.json, test.json (captions)
• Pose: train_pose/json/*.json (keypoints) and train_pose/pose/*.jpg (visualization)
• Sketch: train_sketch/<gt_stem>.png
• Fabric: train_fabric/<gt_stem>.png (fixed-size texture patch)

Data construction pipeline

The mini version follows the EquiFashion-DB construction pipeline:

1. Public sources → raw pool
   Multiple fashion datasets (captioning, recognition, segmentation, editing) are merged into a raw pool with images, captions/attributes, categories and pose/parsing when available.
2. Cleaning & quality filtering
   • Remove broken images, heavy occlusions and extreme truncation.
   • Discard samples with invalid / missing key joints or inconsistent parsing when pose is available.
3. Resolution & category normalization
   • Crop/resize all images to (512 \times 512) around the main garment / person.
   • Map dataset-specific labels into a unified garment taxonomy (40+ categories).
4. Multimodal enrichment (this repo)
   Using the equifashion_pipeline/ code:
   • Generate Canny sketch maps inside garment regions (from pose keypoints when available).
   • Sample high-frequency fabric patches from garment regions.
   • Normalize pose JSON into a unified keypoint format.
5. Packaging
   Final JSON manifests (train.json, test.json, train_pose.json) store standardized paths and captions, with all modalities aligned by filename stem.

References

• [1] Xie et al. — HieraFashDiff: Hierarchical Fashion Design with Mu
• [2] Baldrati et al. — Multimodal Garment Designer: Human-Centric Latent Diffusion Models for Fashion Image Editing (2023)
• [3] Zhang et al. — ARMANI: Part-level Garment-Text Alignment for Unified Cross-Modal Fashion Design (2022)
• [4] Jiang et al. — Text2Human: Text-Driven Controllable Human Image Generation (2022)
• [5] Rostamzadeh et al. — Fashion-Gen: The Generative Fashion Dataset and Challenge (2018)
• [6] Yang et al. — Fashion Captioning: Towards Generating Accurate Descriptions with Semantic Rewards (2020)

Usage

from datasets import load_dataset
ds = load_dataset("NguyenDinhHieu/EquiFashion-DB")

Citation

@dataset{NguyenDinhHieu_EquiFashionDBMini,
  title  = {EquiFashion: Hybrid GAN–Diffusion Balancing Diversity–Fidelity for Fashion Design Generation},
  author = {Nguyen Dinh Hieu, Tran Minh Khuong, Phan Duy Hung},
  year   = {2026},
  url    = {https://huggingface.co/datasets/NguyenDinhHieu/EquiFashion-DB}
}