Yes. ChatGPT can generate new images from a text description and edit an uploaded or previously generated image when image tools are available to your account. You can describe the subject, setting, composition, colors, text, and aspect ratio, then refine the result through conversation. It can create illustrations, concept art, social graphics, product mockups, diagrams, and photo-like scenes. Results still need review for incorrect text, distorted details, unwanted changes, policy limits, and usage rights.
The fastest way to get a useful result is to specify the image’s purpose and what must remain unchanged during edits.
What ChatGPT Images Can Do
OpenAI’s current ChatGPT Images guide describes both generation and editing. Users can request an image in conversation, choose the Images experience, upload an existing image, describe changes, or select an area to edit.
| Task | Useful prompt detail | Review carefully |
|---|---|---|
| Create an illustration | subject, style, palette, composition | anatomy, repeated objects, small details |
| Make a social graphic | platform, aspect ratio, focal point, text | spelling, legibility, safe margins |
| Edit a photo | exact change and protected elements | drift outside the selected area |
| Build a mockup | product shape, viewpoint, lighting, background | logos, labels, realistic proportions |
| Create a diagram | entities, relationships, labels, hierarchy | factual accuracy and text rendering |
| Produce variations | elements that may and may not change | consistency across versions |
Availability can differ by account, product surface, organization, and rollout. Use the controls shown in your account rather than relying on an old screenshot.
For a focused walkthrough of another OpenAI image workflow, see how to use DALL-E to generate images. Product names and interfaces change, but the prompt-planning principles transfer.
How ChatGPT Generates and Edits Images
You begin with natural-language instructions. ChatGPT interprets the request, sends it through the image-generation system, and returns an image. Because the image tool operates within a conversation, your follow-up can refer to the prior result.
OpenAI’s learning documentation for image generation in ChatGPT recommends describing the image and adding a reference image when you want a transformation or visual guidance. It also advises making small, targeted revisions and stating what should remain the same.
Create a new image
Use a prompt with five parts:
- Purpose: where the image will be used.
- Subject: the main person, object, or scene.
- Composition: framing, viewpoint, and placement.
- Visual direction: medium, lighting, palette, and mood.
- Constraints: aspect ratio, text, exclusions, and protected details.
Example:
Create a horizontal editorial illustration for an article about organizing digital notes. Show a calm desk with three clearly separated stacks of cards. Use navy, cream, and coral. Leave open space in the upper-left corner for a title. Do not add words, logos, or extra devices.
This prompt avoids a vague request such as “make an AI productivity image.” It explains the communication job of the picture.
Edit an existing image
Upload or select the image, then name one change. State what must stay the same.
Replace only the blue mug with a small green plant. Preserve the person, desk layout, crop, lighting, colors, and every other object.
The official editor guide warns that selected areas are not always precise and edits may extend beyond the highlight. Compare the whole image with the original after every edit.
Image Types, Customization, and Prompt Controls
ChatGPT can help with a range of visual formats. The right prompt changes with the format.
Illustrations and concept art
Describe medium and visual logic rather than naming a living artist. “Flat editorial vector illustration with geometric shadows” is more controllable than “make it artistic.” Add the audience and emotional tone.
Photo-like images
Specify camera position, scene depth, lighting, and what makes the situation plausible. Review hands, reflections, product labels, and background objects. A polished surface does not prove that the scene depicts a real event.
Graphics with text
Keep the wording short. Put the exact text in quotation marks and say where it should appear. Check every character before publication. If accuracy is critical, generate the visual without text and add typography in a design tool.
Diagrams and educational visuals
Provide the factual structure first. Ask for labels and relationships, then compare the output with a trusted source. Image generation should not be the authority for a medical, legal, scientific, or safety-critical diagram.
Variations and brand consistency
List the elements that define the system: palette, spacing, line weight, viewpoint, and recurring objects. Save a reference image. In each revision, change one variable. This reduces visual drift.
A broader overview of text-to-image workflows can help you choose which details belong in the prompt and which are easier to adjust after generation.
A Step-by-Step Prompting Workflow
Step 1: Write the acceptance criteria
Before prompting, decide what a usable image must contain. For a blog header, criteria might include a wide crop, one focal object, open title space, no embedded text, and a limited palette.
Step 2: Generate a rough direction
Ask for one clear concept. Do not request ten unrelated styles. Review composition before worrying about tiny details.
Step 3: Choose the main revision
Identify the largest gap between the result and the criteria. Correct that gap alone. Examples include moving the subject, reducing clutter, or changing the aspect ratio.
Step 4: Protect successful elements
Name what must remain fixed. “Keep the lighting, viewpoint, and background unchanged” is more useful than “make it better.”
Step 5: Inspect at full size
Check text, hands, edges, shadows, reflections, logos, and repeated objects. Zoom into the area you edited and scan the rest of the image for unintended changes.
Step 6: Save an approved version
Download the selected image and record the prompt, date, and intended use. Keep the original if you edited an existing asset. This makes later corrections easier.
A guide to writing better AI prompts can improve this workflow without tying it to one image tool.
Limitations, Safety, and Rights
OpenAI’s launch material on native image generation describes strengths such as instruction following, text rendering, and conversational refinement. It also lists limitations involving cropping, hallucinations, binding, precise graphing, dense small text, and editing precision. Treat those as review targets.
Common failure modes
- The image contains the right objects but assigns traits to the wrong one.
- Text looks correct at a glance but contains a misspelling.
- A local edit changes lighting or faces elsewhere.
- A product mockup invents labels or controls.
- A diagram is visually neat but factually wrong.
- Several variations lose the original composition.
Safety rules can block some requests or transform them. OpenAI’s usage policies describe prohibited uses and other restrictions. An allowed generation is not proof that you have permission to publish it.
Copyright, trademark, privacy, and publicity rights depend on the source material, jurisdiction, and use. The U.S. Copyright Office’s AI initiative tracks policy and reports about artificial intelligence. For a practical overview, this article on who owns AI-generated images explains the questions to ask before commercial use.
Do not upload an image you lack permission to process. Avoid making deceptive media about real people. Get legal advice for high-risk commercial, political, medical, or identity-related use.
Realistic Use Cases and Quality Checks
Blog illustration
Create an original scene that communicates the article’s idea. Check that the image does not imply a factual event that never occurred. Add alt text based on what the approved image actually shows.
Product concept
Use image generation for early visual exploration, not manufacturing specifications. Label the image as a concept when someone could mistake it for a finished product.
Photo cleanup
Remove a distracting object or change a background. Compare faces, hands, edges, and lighting with the original. For conventional editing techniques, see how to use AI to edit photos.
Presentation graphic
Request a simple visual hierarchy and leave text to slide software. Check contrast, readability, and whether decorative elements compete with the message.
A 15-minute worked example
A small-business owner needs a wide image for a scheduling article. They spend three minutes defining the purpose and acceptance criteria, five minutes generating one composition, four minutes correcting clutter, and three minutes checking edges and text. The process has four review points, not an endless prompt loop. If the image still fails the criteria, they change the concept rather than adding more adjectives.
What to Know Before Deciding: A Decision Framework for Generate, Edit, or Use Another Method
Choose generate when originality and visual exploration matter more than exact factual reproduction. Choose edit when you own or may use the source image and need a bounded change. Use a conventional design, photo, chart, or specialist when precision and verifiability dominate.
Ask these questions:
- Do I have rights to every uploaded reference?
- Could the result be mistaken for a real event or person?
- Must text, measurements, or labels be exact?
- Can a human review the whole image at full size?
- Is the image acceptable under platform and organizational rules?
- Can I explain how it was made if disclosure is expected?
If exactness is essential, generate a draft concept and rebuild the final asset with controlled tools.
Product, Course, App, and Platform Experience
The durable skill is writing visual acceptance criteria. Tool interfaces and model names will change. Purpose, composition, constraints, review, and rights checks will remain useful.
If you want structured practice with image prompting and responsible AI workflows, explore Coursiv AI lessons. Start with low-risk visuals and build a repeatable review habit before using generated images in public work.
A pre-publication image review
Review the final asset in three passes. First, check meaning: does the image communicate the intended idea without a misleading implication? Second, check craft: inspect text, anatomy, edges, shadows, reflections, logos, and repeated objects. Third, check permission: confirm the references, people, trademarks, and intended use are allowed.
Use a second reviewer for public or client work. Do not tell the reviewer which defects you expect. Ask them to describe what the image shows, read every word, and point out anything implausible. A fresh viewer often notices a wrong label or unintended symbol that the prompt writer has stopped seeing.
For a batch, define a rejection rule before generation. Reject any image with unreadable text, a distorted brand mark, an unapproved person, or a factual diagram error. This prevents schedule pressure from lowering the standard after several attempts.
Keep a lightweight creation record
Record the date, tool, prompt, references, major edits, reviewer, and intended channel. The record does not need to expose private conversation history. Its purpose is to make correction and disclosure practical if a question appears later.
When an edit changes a real photograph, keep the original and the approved final separately. Name them clearly. Do not overwrite the source file. If the change could affect the meaning of an event, product, or person’s actions, use a conventional edit or add an appropriate disclosure.
A generated image should enter the same publishing workflow as any other asset. It needs accessibility text, size and crop checks, rights review, and final approval. Fast creation does not remove those steps. It only changes how the first draft was produced.
Frequently asked questions
Can ChatGPT create an image from text?
Yes, when image generation is available in your account. Describe the purpose, subject, composition, style, and constraints, then review the result.
Can ChatGPT edit my existing image?
Yes. Upload or select an image and describe the change. State what must remain unchanged and inspect the whole result because edits can extend beyond a selected area.
Why is the text in my generated image wrong?
Image systems can still misspell or distort text. Keep wording short, verify every character, or add final typography in a design tool.
Can I use a generated image commercially?
That depends on the source material, applicable rights, policies, contract terms, and jurisdiction. Confirm permissions and obtain qualified advice for high-risk use.