Accept the dataset terms
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
Please provide your institutional details and read the following terms
before requesting access.
By requesting access to this food dataset, you agree to the following:
You will not use the dataset to conduct experiments that cause harm
to humans or to generate harmful content.You are solely responsible for your use of the dataset, including
any misuse and any incidents, injuries, losses, or other harm
resulting from your use, to the extent permitted by applicable law.You will comply with all applicable laws, regulations, and ethical
requirements when using the dataset.The dataset authors shall not be liable for any damage, loss, incident, or other adverse consequence
arising from the use or misuse of this dataset.Users are solely responsible for how they use the
dataset and for any resulting consequences.By downloading this dataset, you acknowledge that you have read and agree to comply with all licenses
and terms of use applicable to the dataset and its constituent materials.
Log in or Sign Up to review the conditions and access this dataset content.
FoodClaim:
A Benchmark for Detecting Generated and Manipulated Food Defects
Ahmad Javed1 · Syed Talal Wasim1,2 · Hamid Suleman1,2 · Juergen Gall1,2
1 University of Bonn
2 Lamarr Institute for Machine Learning and Artificial Intelligence
WACV 2027 · Datasets Track
For non-commercial research use only.
Introduction
Generative image models make it easy to fabricate photographic evidence, and food delivery platforms already receive refund claims backed by generated or edited images of spoiled or damaged meals. Detecting such images differs from standard synthetic image detection: a deceptive complaint image is often an authentic photograph with a small edited region, and the question is not only how the image was produced but whether the depicted defect is real. We introduce FoodClaim, a benchmark of 220,066 images across 13 food categories and five classes that separates image authenticity (authentic, generated, or manipulated) from food condition (defect-free or defective). Every generated and manipulated image is derived from a specific authentic photograph with a shared description, defect instruction, and generator family; defects follow a plausibility-rated taxonomy of 78 food-defect combinations, and seven of the 12 generator configurations, including closed commercial models, are held out from training. Across 14 pretrained backbones, a synthetic image detection baseline, and vision-language models, the best detector falls from 88.2% accuracy on seen generators to 63.8% on held-out ones, and frequency-domain inputs do not close the gap. FoodClaim will be publicly released with all descriptions, defect instructions, and source-image lineage.
Dataset
FoodClaim contains 220,066 images across 13 food categories and five classes.
| Class | Train | Test | Total |
|---|---|---|---|
| Authentic | 40,939 | 13,869 | 54,808 |
| Authentic-Defective | 1,006 | 338 | 1,344 |
| Generated | 40,939 | 13,836 | 54,775 |
| Generated-Defective | 40,791 | 13,782 | 54,573 |
| Manipulated-Defective | 40,939 | 13,627 | 54,566 |
| Total | 164,614 | 55,452 | 220,066 |
Food categories: biryani, pizza, burger, french fries, fried chicken, sandwich, noodles, sushi, dosa, dumplings, kebab, salad, and tiramisu.
Each authentic source image is matched with generated, generated-defective, and manipulated-defective variants. Genuine defective photographs are collected separately. Source images and their derivatives stay in the same split.
The val split in the dataset corresponds to the test split in the paper (val = test). The counts above use the paper's terminology.
Sources and generators
Authentic images come from Food-101, Sushi-50, FoodX-251, UECFood256, and Khana, with additional photographs of genuine defects collected from customer complaints.
Training uses FLUX.1-schnell, FLUX.1-dev, FLUX.2-dev, FLUX.2-Klein-9B, and FLUX.2-Klein-4B. Held-out evaluation uses FLUX.2-pro, GPT-Image-1.5, GPT-Image-2, Gemini-2.5-Flash-Image, Gemini-3.1-Flash-Image, and Qwen-Image-2512. GPT-Image-2 is evaluated at two quality settings. FLUX.1-Kontext-dev and Qwen-Image-Edit-2511 provide editing where needed; Qwen3.5-27B produces descriptions.
See the project website for the construction pipeline, examples, category counts, and baseline results.
Examples
Matched examples from the dataset, also shown on the project website. In each row, from left to right: Authentic, Generated, Generated-Defective, and Manipulated-Defective. The variants share the source description, defect instruction, and generator family.
Salad · FLUX.2-dev. Defect: a dead ant partially hidden beneath a lettuce leaf.
French fries · FLUX.2-pro. Defect: a tiny metallic fleck on the oily surface of a fry.
Noodles · FLUX.2-dev. Defect: a thin hair strand tucked between noodles near the bowl's edge.
Metadata
Each sample records its image path, split, authenticity/condition label, food category, generator, source dataset, source-image lineage, structured description, image dimensions, file size, and SHA-256 checksum.
A shortened example from the supplied metadata (the description is excerpted for readability):
{
"schema_version": "1.0",
"sample_id": "0003487d19c79a24e858",
"image_path": "data/val/fake_bad/khana/biryani/0003487d19c79a24e858.jpg",
"split": "val",
"data_category": "fake_bad",
"generator": "gpt-image-1.5-medium",
"source_dataset": "khana",
"class_name": "biryani",
"description": "FOOD:\n- Name: Spinach Rice (likely Palak Pulao)\n- Appearance and color of the dish: A mound of rice grains dyed a distinct, medium-green color. The rice is topped generously with white shreds.\n...",
"width": 1024,
"height": 1024
}
Use data_category for the five-way target and class_name for the food category. The description field contains structured text about the food, serving vessel, background, and image quality.
License and use
FoodClaim is intended for non-commercial research only. Commercial use and operational refund adjudication are outside the proposed permission. Do not use synthetic images as genuine complaint evidence.
See the research-use terms. Original photographs retain their respective rights; model-weight licenses do not automatically license dataset images. FLUX, Gemini, GPT-Image, and Qwen have distinct terms, including output-use restrictions where applicable.
Citation
Please cite FoodClaim if you use it in your research. Proceedings details are preliminary.
@inproceedings{javed2027foodclaim,
title = {FoodClaim: A Benchmark for Detecting Generated and Manipulated Food Defects},
author = {Javed, Ahmad and Wasim, Syed Talal and Suleman, Hamid and Gall, Juergen},
booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
year = {2027}
}
- Downloads last month
- 1,123