Hugging Face adds vision to its AI spreadsheets

Hugging Face has updated AI Sheets so you can work with images inside a spreadsheet without writing code. The tool can now analyze photos, receipts, and documents, extract their data, generate new images, and transform existing ones using open AI models.

AI Sheets works like a spreadsheet in which each column can become an AI action. You can upload images directly or use a dataset that already includes them, write an instruction, and let the model create a new column with the result.

The platform uses thousands of models available through Inference Providers, services that run the models for you. That lets you choose faster options, specialized models, or systems with greater reasoning capacity without installing all the infrastructure needed to run them.

From an image to a useful table

The vision feature can turn visual information into structured data. For example, you can upload a folder of receipts and ask AI Sheets to extract:

• Merchant name
• Purchase date
• Total amount
• Expense category

The result appears in columns that you can review, correct, and export as CSV or Parquet. This can save you hours of manual data entry when you need to organize expenses, inventories, or old documents.

It also works for digitizing handwritten recipes. In a demonstration, Hugging Face uses photos of old recipes to extract ingredients, quantities, cooking times, and other details. The default instruction transcribes all visible text, including headings and advertisements that are not part of the recipe.

When you replace it with a more specific prompt, such as asking for only the ingredients and instructions, the result becomes much more useful. The difference matters: AI does not automatically know which part of an image you care about. The quality also depends on how you frame the task.

You can review and improve the results

Each action combines an instruction and a model. If the response contains an error, you can edit it manually, mark correct results with a thumbs-up, and regenerate the others. Those corrections become examples that help guide subsequent responses.

The tool includes actions for describing and classifying images, detecting objects, extracting text, asking questions about an image, translating content, or adding labels and metadata. In text columns, you can also summarize, extract keywords, or transform information into another format.

The default vision model in the example is Qwen/Qwen2.5-VL-7B-Instruct, chosen for its balance of speed and accuracy. You can also try larger models, such as Qwen/Qwen3-VL-235B-A22B-Reasoning, when an image is harder to interpret. With a complex recipe, both models produced similar results but differed on relevant details, including one ingredient and the cooking temperature.

That points to a basic limitation: you should review an automatic extraction before using it to make decisions or publish information. A model can misread one word even when the rest of the result looks convincing.

Generate images from the same sheet

AI Sheets does not only analyze images. It can also create and edit them from text.

Imagine a sheet with the titles for a social media calendar about healthy recipes. You can add a column with an instruction such as: generate an appetizing photograph of the dish, with natural light and a top-down view. The tool will create an image for each row.

You can then add another action to change the background, add ingredients, or produce different versions of the same image. Editing models such as Qwen-Image-Edit can modify photographs generated or uploaded to the platform.

The result is a complete workflow in one place: text, images, reviews, and export. When the dataset is ready, you can send it to Hugging Face Hub, keep it private, or share it with an organization.

AI Sheets is open source and you can try it without installing anything. You can also run it locally from its GitHub repository. Hugging Face recommends the PRO plan for people who need 20 times more monthly inference usage, meaning more capacity to process queries with AI models.

The important development is not just that a spreadsheet can see images. It is that tasks that once required programming, separate tools, and a lot of manual work can now be organized as columns, instructions, and reviews. Accuracy will still depend on the model you choose, the quality of the image, and human review of the results.