README.md

metadata

pretty_name: CommerceGen-Bench
language:
  - zh
license: cc-by-nc-nd-4.0
task_categories:
  - image-to-image
tags:
  - e-commerce
  - image-editing
  - multi-image-editing
  - visual-text-rendering
  - benchmark
size_categories:
  - n<1K
configs:
  - config_name: single_image
    data_files:
      - split: test
        path: data/single_image_test.parquet
  - config_name: multi_image
    data_files:
      - split: test
        path: data/multi_image_test.parquet

CommerceGen-Bench

CommerceGen-Bench is an e-commerce product-poster generation and editing benchmark containing 400 test cases: 200 single-image editing cases and 200 multi-image editing cases. Each case includes fixed with-PE and without-PE instructions, Chinese target selling-point text, embedded input images, and a three-level product taxonomy.

The release contains test inputs only. It provides neither a training split nor reference output images. Input images are stored directly in the Parquet files, so no separate image directory or public image URL is required.

Files and configurations

• data/single_image_test.parquet: 200 single-image editing cases.
• data/multi_image_test.parquet: 200 multi-image editing cases.
• README.md: dataset card, schema, protocol, limitations, and license.

The single_image and multi_image configurations each expose a test split containing 200 cases. Both prompt conditions share the same input images; there is no duplicated with-PE/without-PE image export.

There is intentionally no default configuration. Load the two tracks separately and concatenate them locally if a combined 400-case dataset is required. Keeping the physical Parquet files assigned to only one configuration avoids duplicated size accounting in the Hugging Face Dataset Viewer.

Schema

Field	Type	Meaning
sample_id	string	Stable anonymized identifier drawn from commercegen_0001 through commercegen_0420. IDs are non-contiguous because 20 cases were removed during curation.
track	string	single_image_editing or multi_image_editing.
product_image	Image	Primary product input. It is the only image input for single-image cases and the first image input for multi-image cases.
logo_image	Image or null	Auxiliary logo/reference input for multi-image cases; null for single-image cases.
instruction_with_pe	string	Frozen enhanced instruction used in the historical with-PE evaluation.
instruction_without_pe	string	Frozen strict no-PE instruction containing the five-requirement prefix, newline-separated selling points, and the strict suffix described below.
target_text_list	list of strings	Selling-point strings expected to appear in the generated poster.
top_level_category	string	One of 6 top-level product groups.
category	string	One of 15 product categories.
subcategory	string	Fine-grained product category. The current release contains 47 non-empty subcategories.

Business identifiers, credentials, signed download URLs, private storage locations, and source filesystem information are not included in this public release. Embedded image fields contain image bytes only and do not expose source file locations.

Dataset composition

Tracks

Track	Cases
Single-image editing	200
Multi-image editing	200
Total	400

Top-level categories

Top-level category	Cases	Share
Beauty & Fashion	157	39.25%
Home & Decor	114	28.50%
3C & Appliances	50	12.50%
Sports & Baby	25	6.25%
Food & Kitchen	45	11.25%
Others	9	2.25%
Total	400	100.00%

Product categories

Category	Cases	Share
Apparel	57	14.25%
Footwear	41	10.25%
Bags & Accessories	28	7.00%
Beauty & Care	31	7.75%
Home Textile	22	5.50%
Lighting	20	5.00%
Hardware	40	10.00%
Furniture	32	8.00%
Digital 3C	27	6.75%
Home Appliances	23	5.75%
Outdoor	18	4.50%
Baby & Toys	7	1.75%
Kitchenware	24	6.00%
Food & Beverage	21	5.25%
Pets & Miscellaneous	9	2.25%
Total	400	100.00%

The taxonomy was assigned according to the actual primary product shown in each case. Category counts are not artificially equalized.

Image and task protocol

Pass product_image first. For a multi-image case, pass logo_image second as a separate image; do not concatenate the two inputs. Select exactly one instruction column for a given experiment and record the prompt condition together with all model settings.

The release does not prescribe a generation resolution. Model-specific input resizing, output size, sampling steps, guidance scale, random seed, and other generation parameters should be recorded separately.

Generated images should preserve the product identity and, for multi-image cases, accurately reproduce the auxiliary logo/reference condition. All requested selling-point text should be rendered correctly and completely according to the selected instruction.

PE and no-PE protocol

instruction_with_pe contains the frozen enhanced instruction generated by Qwen3.8-Max for the experiments reported in the paper. These instructions are released as fixed benchmark inputs and are not regenerated at evaluation time.

instruction_without_pe uses the historical strict C variant: the fixed five-requirement poster prefix, the unnumbered selling points joined with newlines, followed by exactly:

请严格按照我给的卖点内容进行生产,不允许添加新的卖点

The ASCII comma above is preserved exactly as used in the historical evaluation prompt.

The historical B variant used by some open-source runs omits the final suffix. B and C are not identical. To reproduce a documented B run, remove only the final newline and the exact suffix above; do not otherwise rewrite the instruction.

Prompt line breaks are preserved. Comma-compacted prompt variants used in presentation slides are not substituted into the released instruction fields.

Evaluation

The paper evaluates generated images using Gemini 3 Flash with five dimensions:

• Product Fidelity (PF)
• Instruction Following (IF)
• Product Integration (PI)
• Naturalness Detection (ND)
• Artifacts Detection (AD)

Overall is the arithmetic mean of the five dimension scores:

Overall = (PF + IF + PI + ND + AD) / 5

Single-image and multi-image tracks must be reported separately. With-PE and without-PE conditions must also be reported separately.

For reproducible comparisons, record the exact generation model version, inference configuration, random seed, judge model ID, judge decoding configuration, evaluation prompt version, number of repeated judgments, and score-aggregation procedure.

Because commercial poster generation admits multiple valid layouts and visual treatments, this benchmark does not use a single pixel-level reference as the primary evaluation target.

Important reproducibility notes

The historical PE instructions are preserved without silently repairing them:

• 29 cases do not contain at least one exact target selling-point string after removing whitespace differences (31 missing strings in total). This substring check does not prove that the corresponding semantic meaning is absent.
• 67 cases contain an image marker beyond the number of actual inputs, including references to [image3] although cases have at most two input images.
• The historical PE instruction for commercegen_0301 includes a merchant storefront address. This case must be reviewed before public distribution.

Any corrected prompt set should be released as a separately versioned configuration rather than silently replacing the prompts used in historical evaluations.

Load from Hugging Face

Install a recent compatible environment:

python -m pip install -U datasets pyarrow Pillow

Load the two tracks separately:

from datasets import load_dataset, concatenate_datasets

single = load_dataset(
    "TaobaoTmall-AlgorithmProducts/CommerceGen-Bench",
    "single_image",
    split="test",
)

multi = load_dataset(
    "TaobaoTmall-AlgorithmProducts/CommerceGen-Bench",
    "multi_image",
    split="test",
)

assert len(single) == 200
assert len(multi) == 200

dataset = concatenate_datasets([single, multi])
assert len(dataset) == 400

sample = multi[0]
input_images = [sample["product_image"], sample["logo_image"]]
prompt = sample["instruction_with_pe"]

Private repositories require authentication by an account with access.

Load local Parquet files

from datasets import load_dataset, concatenate_datasets

single = load_dataset(
    "parquet",
    data_files={"test": "data/single_image_test.parquet"},
    split="test",
)

multi = load_dataset(
    "parquet",
    data_files={"test": "data/multi_image_test.parquet"},
    split="test",
)

dataset = concatenate_datasets([single, multi])
assert len(dataset) == 400

License

CommerceGen-Bench is released under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0) License.

• The dataset may be copied and redistributed for non-commercial purposes with appropriate attribution.
• Commercial use is prohibited.
• Modified or adapted versions of the dataset may not be publicly distributed.

License terms: https://creativecommons.org/licenses/by-nc-nd/4.0/

Brand names and trademarks remain the property of their respective owners. Use of this dataset must also comply with applicable privacy, publicity, trademark, and other third-party rights.

Citation

@article{xu2026commercegenstudio,
  title={CommerceGen-Studio: An End-to-End Framework for Commercial-Grade Main-Image Generation},
  author={Xu, Xiaoli and Wu, Zhengtao and Xu, Zhengze and others},
  year={2026}
}

Please replace this provisional citation with the full author list and arXiv or conference metadata when available.