How to evaluate result quality in the 'What's on the picture' service
The 'What's on the picture' service helps describe visible image details and answer questions about them, separating observations from inferences. But for the result to be useful, it needs to be critically evaluated. In this guide we will explain how to check accuracy, completeness and suitability of an answer on a specific example. You will learn what parameters to consider, how to formulate a request and what limitations the tool has. We will not promise perfect results: the model may make mistakes, and quality depends on image clarity and your request. Everything you read here is recommendations based on the logic of similar systems, not on fabricated tests. Examples are illustrative in nature.
Open tool: Image description ↗What the tool does and does not do
'What's on the picture' is a web form where you upload an image (up to 4 MB) and enter a text request (up to 8000 characters). The service analyzes the picture and outputs a text description or an answer to your question. It is important to understand: the tool does not generate photorealistic images and does not create audio or video. It works only with visual content and text. If you upload audio, it will not be sent to AI — the service has a separate speech recognition function that runs locally, but this does not apply to image description.
The tool does not perform OCR for spelling check, does not have comparison modes, does not automatically save history and does not publish results directly. You can copy or download the answer in TXT format. Vector formats are only SVG and PNG, speech synthesis is WAV, subtitles are SRT. There is no batch processing, integrations or premium plan. All this limits use cases, but does not prevent getting a useful description if the task is set correctly.
Main feature: the service tries to distinguish observation (what is visible) from inference (interpretation). For example, 'a man in a blue shirt in the photo' is an observation, while 'he looks tired' is already an inference. The quality of the result largely depends on how clearly you ask to separate these categories.
Request: 'List main objects, colors and location in this image. Separately indicate what is an observation and what is an assumption.'
Evaluation criteria: accuracy, completeness, relevance
Accuracy is correspondence of the description to what is really visible in the picture. Check that the tool did not invent extra details: for example, did it name a color of an object that is not there, or did it attribute emotions to a person without basis. If there are statements in the answer that cannot be verified from the image, this is a reason to doubt quality.
Completeness is how much the description covers important elements: main objects, their mutual arrangement, colors, textures, possible text on the image. But completeness does not mean listing everything. A good answer highlights the main points and does not bore with trivialities. If you asked a specific question, completeness is assessed by whether a comprehensive answer was given to it.
Relevance is correspondence of the answer to your request. If you asked about colors but got a plot description, it is irrelevant. The tool may go off track, especially if the request is formulated vaguely. Therefore before evaluation make sure you yourself clearly understand what you wanted to find out.
Image: table with a coffee cup and a laptop. Answer: 'On the table stands a white coffee cup, next to an open laptop. A spoon is visible to the left. Lighting is warm, probably evening.' Here observations (cup, laptop, spoon) are separated from an inference (lighting).
How to formulate a request to facilitate checking
The more specific the request, the easier it is to evaluate the result. Instead of 'What is here?' better write: 'List objects, their colors and mutual arrangement. If you see text, provide it. Do not make assumptions about what is not visible.' Such instruction sets boundaries and reduces the risk of inferences.
It is useful to ask to separate observations and inferences. For example: 'First describe only what is visible, without interpretations. Then, if needed, add inferences but mark them as such.' This will help you quickly distinguish facts from guesses and check accuracy.
If the image is complex, break the task into parts: first ask for a general description, then clarify details. But remember: the service does not support dialogue, each request is processed separately. So it is better to include all requirements in one request immediately. Take into account the 8000 character limit — more than enough for a detailed instruction.
Request: 'Describe what is shown in the picture. Indicate main objects, their colors and location. If there is text, quote it. Do not guess what is not visible. At the end separately list what are your inferences.'
Practical example of evaluation on a specific image
Suppose you uploaded a photo of a room. The tool output: 'On the photo living room with a sofa, coffee table and TV. The sofa is gray, the table is wooden. A picture hangs on the wall. Light from the window on the left. Probably day, as shadows are short.'
We evaluate accuracy: are all objects really there? If there is no TV in the picture, this is an error. Check colors: the sofa may not be gray, but beige. Arrangement: is the window really on the left? Completeness: are important details not missed, e.g. carpet or plant? Relevance: if you asked to describe the interior, the answer corresponds.
Separately evaluate inferences: 'probably day' is an interpretation based on shadows. If shadows are short, it is reasonable, but not absolute. You can accept this as an inference, not as a fact. If the tool did not mark an inference, this is a minus to quality. Overall, the answer can be considered satisfactory if it does not contain gross errors and covers main elements.
Conclusion: 'Observations: sofa, table, picture, window on the left. Colors: sofa gray, table brown. Inference: time of day — day.'
Typical errors and tool limitations
The tool may make mistakes in details: confuse similar objects, incorrectly determine color, skip small text. This is due to model limitations. Also it does not always clearly separate observation and inference if you did not explicitly ask for it. Sometimes the answer may be too general or, conversely, excessively detailed without reason.
The service does not support OCR for spelling check, so if there is text in the image, the tool may quote it but does not guarantee accuracy. There are no comparison modes, so you cannot upload two images and ask to find differences. No batch processing — each image must be uploaded separately.
Size limitations: image up to 4 MB, text up to 8000 characters. If the picture is larger, it must be compressed. This may affect description quality. Also answers may contain errors, and queue and availability of the service vary. Do not rely on the tool in critical tasks without verification.
If there is a sign with text in the photo, the tool may read it with an error. Check against the original.
How to use the result practically
The result can be used as a draft description that you then edit. For example, for writing alternative text for an image, for creating captions or for extracting information. But always verify facts. If you use the description for decision making, make sure critical details are confirmed.
The tool is useful for quickly familiarizing yourself with image content when there is no time to look at it manually. But it does not replace careful looking. If you need exact text recognition, it is better to use specialized OCR services. If deep scene analysis is required, a person may be required.
Remember that the service does not automatically save your data. Copy the result immediately if you need it. Downloading to TXT is a convenient way to save text. Do not upload confidential images if you are not sure about the privacy policy of the service.
Copy the answer to TXT and mark doubtful places so you can check later.
Final check: checklist for evaluation
Make yourself a simple checklist. Accuracy: do all statements correspond to the image? Are there invented objects? Completeness: are main elements covered? Is there an answer to your question? Relevance: did the tool go off track? Separation: are inferences marked as such?
If the answer is unsatisfactory, try reformulating the request, adding more instructions. But do not expect an ideal: the model may make mistakes. If quality is critical, consider alternative methods. The tool is an assistant, not a replacement for expertise.
In general, 'What's on the picture' can be useful for quick descriptions and answering questions if you understand its limitations and are ready to check the result. Use it as a starting point, not as truth in the last instance.
Checklist: 1) Accuracy — yes/no; 2) Completeness — yes/no; 3) Relevance — yes/no; 4) Inferences marked — yes/no.
Evaluating result quality in the 'What's on the picture' service requires a careful approach. Check accuracy, completeness and relevance, separate observations from conclusions. Formulate requests specifically, indicating what needs to be separated into facts and assumptions. Remember the limitations: the tool does not generate images, does not process audio, does not have OCR for spelling check and does not automatically save history. Use it as an auxiliary means, not as a source of absolute truth. Examples in the article are illustrative in nature and are not results of real tests. Always verify important information.
Open tool: Image description ↗Interface translations are machine-assisted and may need refinement.