How to describe images in the tool 'What's in the picture': observation and inference
'What's in the picture' is a web tool that accepts a text instruction and returns a description of visible image details or an answer to a question about it. Its main feature is the focus on distinguishing observation from inference: you can ask separately to list what is actually visible and separately — possible interpretations. The tool does not generate photorealistic images and does not process audio: speech is synthesized via eSpeak, and vector images are saved in SVG or PNG. In this article we will figure out how to get useful descriptions using formulations, check them and avoid typical traps. All examples are fabricated and provided for clarity.
Open tool: Image description ↗How to formulate a request to get an accurate description
The tool accepts a text instruction in the input field (up to 8000 characters) and returns one result in a separate panel. To make the description useful, specify exactly what you are interested in: objects, their color, location, shape, texture or quantity. The more specific the criterion, the more predictable the result. For example, the request 'List main objects and their colors' will give a more structured answer than 'Describe the picture'.
It is useful to explicitly ask to separate observation and inference. The formulation 'First list only what is visible, then separately speculate what it might be' helps separate facts from interpretation. This reduces the risk that the model will output a guess as an observed detail.
Do not mix contradictory requirements in one request, for example 'describe briefly and list all the smallest details'. This confuses the model and leads to an incoherent answer. Better make two separate requests and compare results manually.
Example request: 'List main objects, their colors and their mutual location. Do not make assumptions.'
Different accents: brevity, color, space
The same frame can be described differently, changing the accent in the instruction. A brief description is suitable for quick orientation: 'Table, cup, book, window, tree'. A description with a color accent is useful for design: 'White cup, blue book, green tree outside the window'. A description with spatial relationships helps in navigation: 'The cup is to the left of the book, the window occupies the right part of the frame'.
Each option is obtained as a separate request: the tool does not automatically offer multiple descriptions. You formulate an instruction, press 'Generate' and get one result. To compare options, save them manually — by copying or downloading as TXT.
Such manual enumeration gives flexibility: you yourself decide which aspect is important in a specific task. For example, for accessibility checking a description of location is useful, and for picking a palette — color.
Request 1: 'Describe briefly'. Request 2: 'Describe which colors predominate'. Request 3: 'Describe how objects are located relative to each other'.
How to check the result and distinguish observation from inference
After receiving the description, check if there are statements in it that are not in the image. The model may make mistakes: name an object that is not visible, or assign a color that is not there. To reduce risk, ask separately to list observed facts and separately — possible interpretations.
It is useful to ask a clarifying question: 'What visible details confirm that this is a cup and not a glass?' The model should indicate signs: handle, shape, proportions. If it cannot do this, the conclusion is subject to doubt.
Remember that the tool does not perform OCR to check text and does not have a comparison mode. You compare descriptions manually. Limitations on input data: images up to 4 MB, audio up to 20 MB or 5 minutes, speech up to 3000 characters.
Request: 'List only what is visible. Then separately write what can be inferred.'
How to refine the description for a specific task
If the description is too general, add a criterion: 'Describe surface textures' or 'Indicate the number of objects of each type'. You can ask to list objects in order from left to right or from top to bottom — this makes the answer more structured.
For technical tasks suitable request: 'Describe geometric shapes and their mutual location, using terms: circle, rectangle, triangle'. For everyday — 'Describe what is happening in the picture, as if you were telling a friend'. Different style of vocabulary gives different useful results.
Refinement is an iterative process: got the result, saw what is missing, changed the instruction and generated again. Keep an eye on request length: not more than 8000 characters.
Request: 'Describe objects from left to right, indicate their approximate size (big/medium/small) and color.'
Typical errors and how to avoid them
Too general request, for example 'Describe the picture', gives a superficial result. Specify exactly what you are interested in. A request with contradictions confuses the model. Expecting the tool to offer several options unjustified: each option is a separate request.
Another mistake — expecting the tool to generate photorealistic images. It works with text description and vector graphics (SVG, PNG). Do not ask to paint a picture — this is beyond the capabilities.
Do not expect the model to remember previous requests: each is processed independently, history is not saved. If you need to compare options, save them yourself.
Instead of 'Describe everything' use 'List three main objects and their colors'.
Tool limitations and realistic expectations
The tool does not support batch processing, automatic saving, integrations and publication. Export is possible only in TXT for text, SVG and PNG for vectors, WAV for speech and SRT for subtitles. No DOCX, PDF, settings panel and comparison mode. This affects work organization: save results manually.
The model may make mistakes, and the queue and availability may vary. Do not rely on the tool for critical decisions without verification. For finance, health and law use it only for educational purposes, not as personal advice.
Consider input data limitations: images up to 4 MB, audio up to 20 MB or 5 minutes, speech up to 3000 characters. If the image is larger than 4 MB, compress it before uploading.
If the image is larger than 4 MB, compress it before uploading.
The tool 'What's in the picture' helps get descriptions of visible details and answers to questions, if requests are formulated specifically and observation and inference are separated. Change accents — brevity, color, space, texture — and compare results manually. Check model statements, ask clarifying questions and remember limitations: no batch processing, history and photorealistic generation. Use the tool as an aid, not as a source of truth. Examples in the article are fabricated and provided for training.
Open tool: Image description ↗Interface translations are machine-assisted and may need refinement.