laziji402/Kwai-Ecom-Bench

KwaiMind E-com Bench

This repository contains the standalone E-com benchmark data package for KwaiMind, designed to evaluate image editing models on e-commerce tasks.

📦 Data Availability

The benchmark contains 11 tasks and 1,100 cases, with 1,300 input images. This repository provides the input images, task manifests, and task-specific judge prompts.

🧩 Tasks

Task	Manifest	Cases	Input images
Virtual Try-On	virtual_tryon	100	2
Clothing Detail	clothing_detail	100	1
Clothing Display	clothing_display	100	1
Universal Wearing	universal_wearing	100	2
Pose Change	pose_change	100	1
Background Replace	background_replace	100	1
Outpaint	outpaint	100	1
Product Extract	product_extract	100	1
Text Edit	text_edit	100	1
Tagline Removal	tagline_removal	100	1
Selling Point Display	selling_point_display	100	1
Total		1,100

📊 Evaluation Dimensions

Each task is evaluated on four dimensions. The G and E prefixes denote general and e-commerce-specific dimensions, respectively.

Task	D1	D2	D3	D4
Virtual Try-On	E5 Fit & Silhouette	E3 Texture	G2 Seamless	G3 Physical
Clothing Detail	E1 Product Id.	E3 Texture	G5 Quality	G6 Aesthetics
Clothing Display	E3 Texture	E4 Model	G3 Physical	G6 Aesthetics
Universal Wearing	E1 Product Id.	E4 Model	G2 Seamless	G3 Physical
Pose Change	E4 Model	E5 Fit & Silhouette	G1 Comply	G3 Physical
Background Replace	G1 Comply	E1 Product Id.	E8 Complete	G3 Physical
Outpaint	G4 Preserve	E8 Complete	E1 Product Id.	G2 Seamless
Product Extract	E6 Cutout	E1 Product Id.	G1 Comply	G5 Quality
Text Edit	E2 TextAcc.	E1 Product Id.	G2 Seamless	G4 Preserve
Tagline Removal	G1 Comply	E1 Product Id.	E8 Complete	G2 Seamless
Selling Point Display	E7 SellingPt.	E2 TextAcc.	G1 Comply	G6 Aesthetics

📂 Layout

images/<task>/*
data/*.json
judge/task_judge_prompts.jsonl
DATA_LICENSE

Each manifest uses paths relative to the dataset root. The corresponding input images are included under images/<task>/.

🚀 Usage

Download the dataset while preserving its directory structure:

hf download laziji402/Kwai-Ecom-Bench --repo-type dataset \
  --local-dir /path/to/KwaiMind-Ecom-Bench

Use the dataset with the code repository:

cd /path/to/KwaiMind
bash benchmarks/ecom/run_inference.sh \
  --checkpoint /path/to/checkpoint-80 \
  --data-root /path/to/KwaiMind-Ecom-Bench

Replace the checkpoint and dataset paths with your local paths. Keep the data/, judge/, and images/ directories together under the dataset root.

📄 License

Kwai-Ecom-Bench materials provided in this repository are released under the Creative Commons Attribution–NoDerivatives 4.0 International (CC BY-ND 4.0) license.

• Sharing unmodified materials is permitted for any purpose, including commercial use, subject to the license terms.
• When sharing the licensed materials, provide appropriate attribution and a link to the license.
• Redistribution of unmodified materials is permitted under the license terms. The license does not permit distribution of adapted materials.

See DATA_LICENSE and the full license text for the applicable terms.

The KwaiMind code is licensed separately; see the code repository for its license. Third-party components retain their respective licenses.