Prompting tips for journalists using AI image generators

This article was migrated from an old version of our website in 2025. As a result, it might have some low-quality images or non-functioning links - if there's any issues you'd like to see fixed, get in touch with us at info@journalism.co.uk.

You wouldn't believe the trouble I had the recently trying to get a generative AI image creator to do what I wanted it to do. I wanted a creative image that depicted a line of people representing different emotions in vibrant colours.

If they were not a pain to crop nicely, they were facing all sorts of directions. And then as my collaborator Luba Kassova pointed out, they were not a diverse bunch. Sometimes they were all Caucasian. Some were all men. One was one man in the middle of a group of women. She wrote about this problem a while back.

So I specified: two men, two women. One of each race: white, black, Hispanic and Asian. Facing all in the right hand direction. I still got lots of Caucasians. I still had people facing the wrong way. I sometimes got five or six people. Some didn't have eyes. The problems magnified.

Eventually we settled on one that seemed good enough (but I eventually scrapped it in favour of a more human approach when we relaunched our website recently).

Nevertheless, it left me with a lot of valuable lessons learned, that I wanted to share with you. One of those lessons is that you have to engineer your prompts with different platforms in mind:

Midjourney

Where it thrives:

• Artistic, conceptual visuals: Excels at creating striking, imaginative, and stylised images—ideal for features, opinion pieces, or magazine-style covers.
• Mood & atmosphere: Great for evoking emotion or illustrating abstract concepts.
• Rapid prototyping: Fast for brainstorming or visual storyboarding.

Where it struggles:

• Literal accuracy: Has difficulty with specific group compositions, realistic faces, or precise actions.
• Complex scenes: Multi-person, coordinated actions often yield oddities.

If it’s not working, try:

• Prioritise clarity over complexity: Focus on the most important visual elements; too many details can confuse Midjourney.
• Break down complex scenes: Generate individuals or pairs separately and assemble them in a design tool.
• Leverage style cues: If realism isn’t working, specify an art style or mood—Midjourney excels at stylised visuals.
• Draw on community wisdom: Search Discord or prompt libraries for examples that match your needs.

DALL-E (OpenAI)

Where it thrives:

• Versatility: Handles a wide range of styles, from photorealistic to cartoonish—good for explainer graphics or spot illustrations.
• Simple scenes: Performs well with single subjects or uncomplicated settings.
• Editing capabilities: Inpainting/outpainting for tweaking or extending images.

Where it struggles:

• Complex group dynamics: Multiple people or coordinated actions can result in errors.
• Text in images: Still unreliable for generating readable, accurate text.

If it’s not working, try:

• Be direct and explicit: Spell out numbers, roles, and actions in your prompt.
• Iterate with variations: Use the “variations” feature to refine results, especially for group scenes or specific details.
• Edit in stages: Use inpainting/outpainting to fix or add elements after the initial generation.
• Try alternative wording: If a prompt isn’t working, rephrase or use synonyms—DALL-E sometimes latches onto different terms.

Magic Media (Canva)

Where it thrives:

• Quick editorial graphics: Perfect for social media, blog headers, or infographics where speed and brand consistency matter.
• Integration: Fits into Canva’s design workflow for easy editing and layout.

Where it struggles:

• Detailed group scenes: Balancing gender, ethnicity, and direction is challenging; anatomy can be inconsistent.
• Complex interactions: Struggles with nuanced actions or multiple people interacting naturally.

If it’s not working, try:

• Keep prompts simple and front-loaded: Canva’s Magic Media prioritises the first details in your prompt and may ignore later instructions.
• Limit variables: Focus on one or two key details rather than complex scenarios.
• Generate separately: For group shots, create individuals and assemble them in Canva.
• Use Canva’s editing tools: Crop, reposition, or combine elements to fix composition or diversity issues.
• Regenerate if needed: Sometimes, simply trying the same prompt again yields a better result.

15 essential tasks GPTs can do for journalists

No, AI cannot replace original journalism. But it can remove some of the everyday tedium so you can focus on what you do best

Journalism UKMarcela Kunova

Adobe Firefly

Where it thrives:

• Brand-safe, commercial use: Trained on licensed content, making it ideal for editorial and commercial projects.
• Style consistency: Good at matching specific visual styles.
• Integration with Adobe suite: Easy to refine images in Photoshop or Illustrator.

Where it struggles:

• Hyper-specific editorial scenes: Large groups or subtle interactions may lack accuracy.

If it’s not working, try:

• Start with the main subject: Generate the key person or object first, then add context or background in post.
• Switch styles: If one style isn’t working (e.g., photo vs. illustration), try another for better results.
• Refine in Adobe suite: Use Photoshop or Illustrator to correct errors or enhance the image.
• Stay updated: Firefly is evolving—new features may address persistent issues.

Stable Diffusion

Where it thrives:

• Customisation: Open source and highly adaptable—can be fine-tuned for newsroom-specific needs or styles.
• Community models: Access to a wide range of pre-trained models for different aesthetics.
• Control: Negative prompts and advanced settings for more precise outputs.

Where it struggles:

• Technical complexity: Requires more setup and prompt expertise.
• Multi-person scenes: Consistency in faces, hands, and group arrangements can be unreliable.

If it’s not working, try:

• Use negative prompts: Exclude unwanted features (e.g., “no extra fingers, no text”).
• Try different models: Some community-trained models handle specific subjects or diversity better.
• Reorder and simplify: The order and simplicity of your prompt can affect results—experiment with phrasing.
• Composite in post: Generate backgrounds and subjects separately, then combine them in an editing tool.

General guidance

Leverage strengths: Use each tool for what it does best—creative, simple, or brand-safe visuals.

Tailor prompts: Be specific, iterate, and combine or edit outputs as needed for the best result.

Address diversity bias: All major tools can reflect bias in their outputs. Always specify diversity in your prompts (gender, ethnicity, age, ability, etc.), and critically review images to avoid stereotypes or exclusion. If needed, generate multiple images and select or edit for fairer representation.

Editorial integrity and transparency: Always label AI-generated images clearly (using captions, watermarks, or metadata) and disclose their use to your audience. Develop and follow newsroom policies for transparency around AI-generated content.

Legal and copyright risks: Be aware that copyright law for AI-generated images is unsettled, with ongoing lawsuits (especially involving Midjourney and Stable Diffusion). Ownership and commercial use rights may be unclear—always check platform terms and seek legal advice before using AI images commercially, except where platforms like Adobe Firefly explicitly guarantee safe use.

Misinformation and deepfakes: Photorealistic AI images can easily be mistaken for real events. Never use AI-generated visuals for crime scenes, political coverage, disasters, or to depict real people or events—especially where accuracy is critical or harm could result. Always fact-check AI outputs for plausibility, anachronisms, or misleading details.

Ethical considerations: Consider the environmental impact of AI image generation (high energy use) and the effect on human creators. Avoid using AI to mimic the style of living artists without consent, and support freelance photographers and illustrators where possible.

Know when to use traditional methods: For sensitive, factual, or high-stakes reporting, traditional photography or illustration is often the safer and more ethical choice. Use AI-generated images mainly for conceptual, illustrative, or background purposes.

Fact-check AI outputs: AI can generate plausible but incorrect or impossible visuals. Always review images for accuracy and context before publication

Note: As these tools evolve, expect their strengths to broaden and their weaknesses to narrow—especially in group composition, realism, and representation. For now, strategic prompt engineering, post-editing, and a critical eye for diversity are key to unlocking their full potential for journalism.

This article was drafted by an AI assistant before it was edited by a human. It was originally published on 2 October 2025 and was updated on 5 November with new information