What's in the picture: how to correctly describe images and avoid mistakes
The "What's in the picture" tool helps describe visible image details and answer questions, distinguishing observation from inference. It works in a web form: you enter a text instruction, select the output language, press the Generate button and get one result in a separate panel. The result can be copied or downloaded as plain text TXT. In this article we will explain how to correctly set tasks to get useful descriptions, and what common mistakes lead to inaccurate answers. All examples in the article are fictional and provided only for illustration.
Open tool: Image description ↗How to formulate a request to the tool
The tool does not guess your goals: it answers what you wrote. If the request is too general, for example "Describe the picture", the result may turn out to be either superficial or overloaded with details you do not need. To get a useful answer, specify exactly what interests you: objects, their color, location, text on the image, relative position of objects or specific details.
It is useful to decide in advance whether you need only observations or also assumptions. The tool strives to distinguish these categories, so you can explicitly ask: "Describe only what is visible, without assumptions" or "Make an assumption about the purpose of the object, but mark it as an assumption". The clearer the instruction, the more predictable the result.
Avoid ambiguous formulations. The phrase "Describe what is wrong" can be understood as searching for errors or as a request to describe unusual details. Better clarify: "Are there objects on the image that look unusual or inappropriate? If yes, describe them and indicate what exactly seems strange."
Instead of "What is here?" write: "List main objects, their colors and relative position. Separately list all visible text. If you need only observations, add: "Without assumptions about what is happening off-screen".
Image quality and size limitations
The accuracy of the description directly depends on the quality of the original image. Blurry, dark, overexposed or heavily noisy images lead to errors: small text may not be recognized, colors are distorted, and object boundaries are indistinguishable. The tool does not increase resolution and does not improve quality, so before loading make sure the image is clear and well lit.
If you need to read text, check that it is readable. If the image is too large, you can crop it, leaving only the important part. This is especially useful when there are many secondary details on the picture distracting from the main one. Remember that the tool accepts images up to 4 MB; if the file is larger, it needs to be compressed or reduced.
Do not upload formats that the tool does not support. It works with images, not documents. If you try to pass a PDF or Word document, the result will not be obtained. Also do not expect handwritten text to be recognized as well as printed: handwriting may be illegible.
If you are photographing a document, make sure the text is clear and not blurry. If necessary, crop the image to the area with text and repeat the request: "Read all visible text on the image".
Observation and inference: how not to get confused
The tool distinguishes between observation (what is visible) and inference (what can be assumed). Users often expect the tool to determine emotions, intentions or purpose of objects, but such judgments always remain assumptions and can be erroneous. If you want to get only facts, explicitly specify this in the request.
For example, in response to the question "Describe what people are doing in the photo" the tool may answer: "People are sitting at a table" (observation) and "probably they are having lunch" (inference). If you need only observations, clarify: "Describe only visible actions, without assumptions about what they are doing". If inference is still needed, ask to mark it as an assumption.
If the tool gave an inference that seems unfounded, check if the request was ambiguous. Perhaps you implicitly asked to make an assumption. In such case reformulate the request and try again. Remember that the model can make mistakes, so critically evaluate the result.
Request: "Describe only visible details. If you make an assumption, mark it with the word 'probably'". Answer: "A cup on the table, probably with coffee".
Tool limitations and typical expectations
The tool is not a universal solver of image-related tasks. It does not perform OCR in proofreading mode, does not generate photorealistic images, does not create audio or video, does not support batch processing, automatic saving, integrations and other functions. If you expect these capabilities from it, you will be disappointed.
There are also technical limitations: the first text request - up to 8000 characters, additional details - up to 1000 characters, images - up to 4 MB. If the limits are exceeded, the tool may not accept the request or truncate it. The tool gives one result in a separate panel, which can be copied or downloaded in TXT format. It is not intended for multiple generation variants or comparison.
Do not use the tool for medical diagnosis, legal conclusions or financial consultations. It describes visible details and answers questions, but does not replace professional judgment. If your task goes beyond image description, it is better to turn to specialized tools.
Do not ask the tool to diagnose by photo. It can describe visible elements, but cannot give a medical conclusion.
How to check and improve the result
Even with an exact request the answer may contain errors. The model can make mistakes in details, especially if the image is complex or ambiguous. Always check the answer, comparing it with the image itself. If the tool claims something you do not see, it may be an error. In such cases reformulate the request and try again.
If the result is unsatisfactory, check if the request corresponds to the tool's capabilities. Perhaps you are asking for something it cannot do. Clarify the request: add specifics, specify exactly what interests you, and ask to separate observations and inferences. Also check the image quality: if it is blurry or dark, try to improve it or choose another frame.
Remember that the tool can give different results with different formulations. Experiment with requests until you get a useful answer. Do not be afraid to ask clarifying questions if the tool supports repeated requests. This will help better understand how it interprets your instructions.
If the tool did not recognize the text, try increasing the contrast of the image or cropping it to the area with text. Then repeat the request: "Read text on the image".
Practical use scenarios
The tool is convenient when you need to quickly get an image description for further work: for example, compose a photo caption, check if all objects are visible in the frame, or extract text. In each case it is important to adapt the request to the specific task. For a caption a brief description is enough, for checking - listing details, for text extraction - exact instruction to read all visible text line by line.
If you work with multiple images, remember that the tool does not support batch processing. You have to upload them one by one and run generation separately for each. This is worth considering when planning time.
Do not forget that model answers may contain errors, and queue and availability may change. If the result seems doubtful, it is better to recheck it manually or repeat the request with clarifications.
For photo caption: "Describe main object, its color and background in one sentence". For checking: "List all objects in the image and indicate if there are people". For text: "Read all visible text and list it line by line".
The "What's in the picture" tool is a useful means for describing images, but its effectiveness depends on the quality of the original data and the accuracy of the request. Avoid vague formulations, check images for clarity, distinguish observations and inferences, consider tool limitations and critically evaluate results. Following these recommendations, you will be able to obtain more accurate and useful descriptions. Remember that the tool does not replace human judgment and does not guarantee accuracy. The examples in the article are fictional and provided for illustration purposes.
Open tool: Image description ↗Interface translations are machine-assisted and may need refinement.