Food photography is documented in many different ways: recipe images paired with cooking instructions, benchmark datasets built for image recognition, photographs taken during real meals, and social posts shared by adolescents and influencers. The numbers below show the scale and variety of that visual record, while keeping each study’s date, geography, and purpose explicit.
Contents
- Recipe-linked image collections
- Fine-grained categories and annotations
- Photographs from real meals
- Where adolescents encounter food images
- Food type, portions, and branding
- Influencers and social contexts
- Restaurant photos and benchmark coverage
Recipe-linked image collections
Some of the largest food-photography collections are connected to recipes. That link matters for home cooks because an image can be interpreted alongside ingredients, preparation steps, and a finished dish rather than as an isolated picture.
The 2018 Recipe1M+ dataset, described in Recipe1M+: A Dataset for Learning Cross-Modal Embeddings for Cooking Recipes and Food Images, contains more than 1 million cooking recipes paired with 13 million food images. Its scale makes it one of the clearest examples of the relationship between written cooking instructions and visual references.
The earlier 2017 Recipe1M corpus, described in Learning Cross-Modal Embeddings for Cooking Recipes and Food Images, contains more than 1 million cooking recipes and 800,000 food images. These counts describe different dataset versions and should not be treated as a single combined total.
Other collections are narrower in geography or purpose. The 2024 FoodLens study curated 30,000 food images from restaurant menus in India and covered more than 140 popular Indian dishes. The collection therefore adds regional and menu-specific coverage that a general recipe dataset may not provide.
The 2019 FoodX-251 dataset contains 158,000 web-collected food images organized into 251 fine-grained food categories. The 2014 dataset release documented by TensorFlow Datasets lists Food-101 at 101,000 real-world food images across 101 food categories.
| Collection | Measurement period | Images | Categories or paired content |
|---|---|---|---|
| Recipe1M+ | 2018 | 13 million | More than 1 million recipes |
| FoodX-251 | 2019 | 158,000 | 251 categories |
| Food-101 | 2014 dataset release | 101,000 | 101 categories |
| FoodLens | 2024 | 30,000 | More than 140 Indian dishes |
These figures are benchmark contents, not estimates of the commercial food-photography market. They also reflect different collection methods: recipe-linked images, web-collected images, real-world benchmark photos, and restaurant-menu images.
Fine-grained categories and annotations
Counting images is only part of understanding a food-photography collection. Labels and segmentation annotations determine what researchers can ask of the images: whether a model can distinguish similar dishes, identify the food area, or work with more than one item in a scene.
FoodX-251 assigns 118,000 images to its training split. It provides human-verified labels for 40,000 validation and test images, including 12,000 validation images and 28,000 test images. The split sizes show how a large collection can reserve a substantial portion for evaluation rather than using every image for training.
The 2021 UEC-FoodPix Complete benchmark, described in A Large-Scale Benchmark for Food Image Segmentation, contains 10,000 images across 102 classes with controlled segmentation annotations. A related 2020 UEC-FoodPix version also contains 10,000 images across 102 classes, but uses automatically generated GrabCut annotations. The annotation method is therefore an important distinction even when the headline image and class counts match.
The same 2021 benchmark paper reports that the 2019 SUEC Food Dataset contains 28,897 images across 256 classes with automatic GrabCut annotations. It also reports that the 2017 Google Food-201 segmentation set contains 12,093 images across 208 classes with crowdworker annotations.
Together, these collections illustrate why “food photography” can mean more than attractive finished-dish images. Some datasets emphasize category recognition, while others describe the boundaries of food within a photograph. For a home-cooking reader, that distinction helps explain why a dataset may contain many visually similar dishes or images where the plate occupies only part of the frame.
Photographs from real meals
Recipe and benchmark collections are useful, but photographs taken during eating capture a different moment: what people actually served, photographed before eating, photographed afterward, or documented from packaged-food labels.
The 2021–2022 SNAPMe benchmark contains 3,311 unique food photos linked to 275 ASA24 food records. It includes photographs from 95 participants who completed at least one study day, and 275 diet days with parallel food photographs and records.
Within that collection, there are 1,475 before-eating photos of non-packaged foods and 1,436 after-eating photos of non-packaged foods. Packaged foods are represented by 203 front-label photos and 196 ingredient-label photos. The benchmark also records that 90 participants completed all three days of data collection.
SNAPMe is not a measure of all meal photography. It is a U.S. meal-photo benchmark from 2021–2022 designed for dietary-assessment computer-vision evaluation. Its value is the pairing of images with food records and the inclusion of before-and-after views, not a claim that these are typical proportions for every household.
Where adolescents encounter food images
The 2018 Flemish adolescent diary study, What Do Adolescents See on Social Media?, collected 638 food-related screenshots over one week. After cleaning, 612 food images remained. The percentages below describe this cleaned sample, not all internet food photography.
Instagram accounted for 69% of shared food images, or 421 images. Facebook or Messenger accounted for 18%, or 110 images, while Snapchat accounted for 9%, or 54 images. YouTube contributed 3%, or 20 images, and WhatsApp contributed 3 food images.
Only 5 of the 612 cleaned images had originally been posted by the participating adolescents themselves. In this study, most of the images adolescents encountered were therefore content created or posted by other people, rather than photographs they had uploaded themselves.
The platform distribution is compact but informative:
| Platform | Images in cleaned sample | Share reported by study |
|---|---|---|
| 421 | 69% | |
| Facebook or Messenger | 110 | 18% |
| Snapchat | 54 | 9% |
| YouTube | 20 | 3% |
| 3 | Not reported as a percentage |
Because the study covered Flemish adolescents during one week in 2018, its platform mix should be read as a dated, population-specific snapshot. It is useful for understanding exposure in that sample, not for forecasting current platform usage.
Food type, portions, and branding
The same Flemish study provides detail about what appeared in the images. Food appeared in 589 of the 612 cleaned images. About 67% of the images, or 409 posts, showed only non-core foods. Core food appeared alone in 137 images, or 22% of the total sample. Mixed core and non-core foods appeared in 58 images, about 10% of the sample.
Portion descriptions were also recorded. Among the 409 non-core-food images, 233 showed excessive quantities. Among the 137 core-food images, 96 showed regular quantities and 39 showed excessive quantities. These are study classifications, not nutritional recommendations or a universal definition of an appropriate serving.
Branding was common in the sample. Branded food products appeared in 289 images, or 47% of the sample. More than 43% of branded food messages were clearly presented as paid marketing, totaling 126 images. The figures describe the cleaned images from the 2018 Flemish diary study and should not be generalized to every social platform or age group.
For readers sharing cooking photos, the distinction between a dish, a portion, a product label, and a paid message is useful. A photograph can communicate food type and quantity while also carrying branding or promotional context.
Influencers and social contexts
Food images often show more than food. In the Flemish adolescent sample, food was linked to a social context in 296 images, or 48% of the cleaned sample. Images associated with spending time with friends numbered 119, and restaurant eating contexts appeared in 100 images.
The study also separated everyday influencers from recognized social-media influencers and traditional-media celebrities. Everyday influencers, mainly peers, posted 164 images, about 27% of the sample. Recognized social-media influencers appeared in 86 images, about 15% of the food images. Traditional-media celebrities were responsible for 48 images, about 8% of the sample.
The participant-level distribution was uneven. Female participants shared 71 influencer images, compared with 15 shared by male participants. Of the celebrity-image shares, 80% came from female participants and 20% from male participants. These figures describe sharing patterns in the study population; they do not establish general differences for all adolescents or all food photography.
Restaurant photos and benchmark coverage
Crowdsourced restaurant photographs can help describe what diners see, but they do not automatically cover every menu item. A 2023 restaurant-image validation study found that crowdsourced food images matched 70.1% of participants’ favorite restaurant foods. Only 44% of restaurant menu items appeared in the crowdsourced food-image data.
Coverage varied by restaurant type. The menu-item match rate was 40% for full-service restaurants and 48% for limited-service restaurants. The correlation between restaurant image counts and foot traffic was 0.24 for full-service restaurants and 0.11 for limited-service restaurants. These correlations are associations reported by the study, not evidence that image counts cause visits.
Another reference point is the 2015 UNICT-FD1200 dataset. It contains 4,754 photos covering 1,200 distinct dishes. Each plate was photographed multiple times, averaging four captures per plate, and the images are labeled across 8 meal categories.
The restaurant-focused figures show both the usefulness and limits of crowdsourced photography. Images may reflect foods diners choose to photograph, while menu items that are less visually distinctive or less frequently shared can remain absent. Benchmark datasets such as UNICT-FD1200 address recognition and repeated views at a defined scale; restaurant validation studies address how well shared images represent available food.