Early ChatGPT Vision Examples: From Diagrams to Code and Visual Mistakes

In October 2023, I collect early examples of ChatGPT's image-input capability while waiting for access myself. [OpenAI began rolling out image and voice features](https://openai.com/index/chatgpt-can-now-see-hear-and-speak/) in September 2023. These selected social-media demonstrations show possibilities, not a systematic accuracy evaluation.

Pietro Schirano shares a four-panel cartoon about people who think they agree but imagine different shapes. ChatGPT interprets the panels as a lesson about communication and reaching shared understanding. I am impressed that its response goes beyond listing visible objects, although an interpretation is not necessarily the only possible reading.

McKay Wrigley shows a cell diagram being explained for a student, with a list of organelles and their functions. I imagine learners uploading textbook illustrations and asking for simpler or more detailed explanations. A generated description still needs to be checked against reliable educational material.

In a food example, Brian Roemmele asks for a recipe from a photograph of a dish. The model supplies a plausible recipe, which could serve as a starting point; a photo alone cannot establish the exact ingredients or preparation used by the original cook.

Aadit Sheth shares an electronics schematic and asks what it shows. The response describes components such as a microcontroller, oscillator, capacitors and USB connector. The example suggests a way to explore a circuit, but a generated explanation is not independently verified by this demonstration.

Schirano uploads a mushroom image and asks for a simulated role-play. The model produces an imaginative response. Neither the scene nor its text identifies the mushroom reliably or demonstrates its effects; the example is role-play, not a test of species or safety.

Another photograph shows a living room, and ChatGPT proposes changes to lighting, color, plants, art and storage. I see potential for design brainstorming. The suggested personal background in the response is not established by the photograph alone.

A detailed brain diagram prompts a broad summary of regions, pathways and senses. The answer is relatively high level, and I suggest asking follow-up questions. An image explanation can sound confident while missing or inventing details, so a reader should verify technical content.

Peter Yang's example asks whether parking is allowed at 4 p.m. on a Wednesday given several signs. The model answers that one hour is allowed. The video does not independently verify the full local rules or all conditions; a model's answer should not replace checking the signs and applicable restrictions.

A photographed page of textbook problems receives step-by-step solutions. I did not check the answers, and the page shown contains physics questions, so the demonstration is not proof that all problems were read or solved correctly. It does show how a student might ask for an explanation of material in an image.

Two interior images lead the model to propose a name, “Athenian modernism,” for a blend of classical features and modern furnishings. That is a creative label suggested for the images, not an established architectural classification.

A user uploads a research-paper image and asks for a high-school-level explanation. The response summarizes the subject of instruction-data selection. A clear answer does not by itself prove that every small line was accurately read or that the summary captures the entire paper.

A separate voice example shows two ChatGPT voices conversing. It illustrates a voice interface rather than image understanding, and the casual conversation is a demonstration of generated dialogue rather than evidence that the systems understand one another as people do.

One user shows a rough house image and asks how to build it. The model begins offering supplies, including a named retailer; the complete materials list is not visible in the video. I speculate that product mentions might create advertising opportunities, but the example does not establish any paid placement or sponsorship.

A crossword image and its clues produce an apparent solution. The original poster and I did not check the answers, so the result should be treated as an unverified example rather than evidence that the puzzle was correctly solved.

A dashboard screenshot is used as the basis for generated web code. The poster then runs the result, suggesting how an image can provide a starting point for implementation. The clip does not establish that the generated site reproduces all functions, is secure or is ready for production.

Another example converts a Figma design into a React component using Material UI. The poster says the model also inferred likely data props. I see potential for shortening design-to-code work, though a visually similar component still requires testing, accessibility checks and integration.

Schirano tests an artwork with a logo embedded in the scene. ChatGPT first describes the landscape without mentioning the logo and then says it sees none; after the user explicitly names the OpenAI logo, it agrees. A leading prompt may influence an answer, so the later agreement alone does not prove independent recognition.

A handwritten home-page sketch becomes a simple page with a heading, input and button. I imagine this making prototyping accessible to nontechnical people. A basic mockup is a starting point, not automatically a complete working business website.

Another shared example asks about an image from the film Her using custom instructions and reports an answer identifying the character and actor. The video presents a historical example; it does not establish that such identification is reliable or that any particular behavior is available today.

Ethan Mollick shares a poker image with two fours in the player's hand and two on the board. The model calls it three of a kind, but the visible four fours make four of a kind. It is a clear example of a plausible answer that miscounts visual evidence; a betting decision would also depend on game context.

An image of two toy figures produces descriptions of their apparent equipment. Berman notes that some details are approximate. The example is a reminder that object recognition and a confident explanation are not the same as a verified identification of each object.

A photographed day planner is used to request a Python Tkinter interface with similar sections for priorities, tasks and notes. The resulting interface is a useful illustration of image-to-code prototyping, subject to the same need to run and test generated software.

A trading-chart example asks for the next trade. The response declines to give a specific trade while describing elements of the chart, according to the video. That behavior in one example does not establish that every chart interpretation or future response will be accurate.

I close by saying I have not yet tested vision myself and hope to investigate it once access arrives. These early user examples suggest a wide range of image-based questions, while the unchecked answers and the poker mistake show why demonstrations should be read alongside their limitations.