AI photography marketing is not a trend you can safely ignore anymore — and it’s not one you can blindly adopt either. At Choco Media, we’ve spent the better part of the last year testing AI-generated imagery in real client campaigns: product shots, lifestyle scenes, social ads, hero images. What we found is a much messier picture than the breathless coverage suggests. Some use cases genuinely saved us hours and produced results that would have cost thousands from a commercial photographer. Others looked fine at 200px preview and completely fell apart in production. This post is our honest field report.
It’s aimed at marketing teams and brand managers who are considering AI imagery tools — Midjourney, DALL-E 3 (via ChatGPT), Adobe Firefly, Stable Diffusion — and want a realistic picture of where these tools earn their place and where they quietly sabotage your brand. We’re not going to tell you AI will replace photographers. We’re also not going to tell you it’s useless. We’re going to tell you exactly which situations call for which approach.
By the end, you’ll have a decision framework for when to use AI-generated photography, when to commission real shoots, and what to watch for at every stage of the workflow.
What AI-generated photography actually is (and isn’t)
Let’s anchor the conversation. When we say “AI photography,” we mean images generated entirely by diffusion models — no camera, no photographer, no stock library. The image is synthesised from a text prompt (and sometimes a reference image) by a model that has learned patterns from billions of real photographs.
This is different from:
- AI-enhanced photography — real photos edited with AI tools (Adobe Generative Fill, Luminar, etc.)
- AI upscaling or retouching — real photos processed through AI to increase resolution or remove blemishes
- Stock photography — even when licensed from platforms that generate AI images on demand
The distinction matters because each category has different strengths, different failure modes, and different legal considerations. In this post we’re focused specifically on fully generated imagery — prompting a tool like Midjourney and using the output in marketing material.
The tools we’ve actually used
Our working stack has shifted over the last year. We’ve spent the most time with Midjourney v6 (and now v7 alpha), DALL-E 3 via the ChatGPT interface, and Adobe Firefly for anything that needs to stay within Adobe’s commercial rights framework. We’ve also experimented with Ideogram for typography-heavy images and Runway for short video clips from stills. Each has a distinct character — Midjourney tends toward a certain cinematic polish, Firefly is safer for commercial use and integrates cleanly with Photoshop, DALL-E is the fastest for iteration.
Where AI photography genuinely works for marketing
Start with the use cases that have proven reliable across multiple client contexts. These are the areas where we now default to AI without a second thought.
Concept visualisation and mood boards
Before a brand shoots anything, someone has to describe what the shoot should look like. AI has almost entirely replaced stock-photo mood boards in our process. Instead of hunting Unsplash for “something close,” we generate a dozen variations of the exact scene, lighting style, and tone we’re after. The client sees the precise direction, not a rough approximation. We’ve found this cuts pre-production rounds by roughly half — fewer “that’s not quite what we meant” moments when you can show the thing rather than describe it.
- Typical time saving: 2–4 hours per concept deck
- Quality bar needed: low — these are internal reference images, not deliverables
- Tool we use: Midjourney (fast iteration), DALL-E (when we want to stay in ChatGPT workflow)
Abstract, atmospheric, and non-literal imagery
Blog post headers, podcast cover art, social backgrounds, email hero images — any context where the image is supporting the text rather than depicting a specific reality. AI is excellent here. You can generate an image that feels like “the texture of a winter morning in a Nordic city” or “the visual weight of a complex decision” without those concepts needing to literally exist as a photograph. These images perform well because they’re more atmospheric than literal stock photography, and they can’t be spotted on a reverse image search running across your competitors’ content.
Scalable social content at volume
If you’re running a content-heavy social strategy — multiple posts per week across channels — AI can generate a consistent visual library at a pace no photography budget can match. The key is establishing a prompt template that encodes your brand’s lighting, colour palette, and compositional style, then varying the subject. We’ve done this for seasonal campaigns where a client needed 60+ unique images over six weeks with a coherent look. A traditional shoot would have cost €8,000–15,000. AI got us there for the price of a Midjourney subscription and about 12 hours of prompt refinement.
“The honest metric is not image quality in isolation — it’s image quality per euro at the volume your strategy actually requires. At scale, AI wins almost every time, as long as the use case is right.”
Paid ad creative testing
Paid media teams need volume and variety. For paid media campaigns, AI lets you test 20 creative concepts at the cost of testing 3 with traditional photography. You find your winners faster, then invest the real photography budget in scaling what works. This is the use case we recommend most to brands running Meta or Google performance campaigns — use AI for the discovery phase, production shoots for scaling winning creatives.
- Test 15–20 AI-generated hero images at €10–50/day per variant
- Identify the 2–3 that drive the highest CTR and lowest CPM
- Commission real photography for those specific scenes
- Scale with confidence that the creative direction is proven
Where AI photography fails (and why it matters)
This is the section that the tool vendors don’t put in their landing pages. Every use case below has burned us at least once, and we’ve seen clients make expensive mistakes by not knowing these limitations upfront.
Product photography with your actual product
AI cannot photograph a product it hasn’t seen. You can get a convincing image of “a minimalist Scandinavian skincare bottle on a marble surface” — but it won’t be your bottle. The logo will be wrong, the cap will be wrong, the exact shade of the packaging will be off. For brands where product accuracy matters (which is almost every e-commerce brand), AI-generated product shots require so much post-editing correction that you’re better off with a light-table and a DSLR.
- The exception: AI-enhanced photography, where you photograph the real product and use AI to change the background, lighting environment, or setting.
- Tools that help: Adobe Firefly’s Generative Fill, Photoroom, Bria — these composite your real product into AI-generated scenes.
Specific people, faces, and identifiable individuals
AI is unreliable with faces. Even with reference images, facial likeness varies unpredictably. More importantly, using AI to generate images that appear to show real, identifiable people — customers, staff, public figures — without their consent creates legal exposure under the EU AI Act and GDPR-adjacent frameworks. We do not use AI to generate imagery of people in client campaigns, full stop. If a campaign needs human faces, we either use real photography with model releases or work with stock libraries where model releases are pre-cleared.
Anything requiring documentary credibility
Journalism, testimonials, case studies, “behind the scenes” — any context where the image implies it depicts something that actually happened. AI imagery used in these contexts isn’t just legally problematic, it’s a trust problem. If a reader reverse-image-searches your case study hero image and gets an AI-generation result, the damage is disproportionate to whatever you gained by avoiding a real shoot.
High-detail technical imagery
Machinery, medical devices, architecture, food with texture — anything where a specialist or an informed reader can spot anatomical impossibility or structural implausibility. AI models are trained on aggregate image data; they know what “a coffee machine” looks like on average, but they don’t know what your specific industrial espresso unit looks like, and they routinely generate machines with wrong valve counts, impossible button placements, and handles attached to nowhere. These errors are invisible at glance and obvious on inspection. That’s the worst possible combination for marketing credibility.
The legal and ethical layer you need to understand
IP ownership of AI-generated images remains genuinely unresolved in most jurisdictions. The current position in the EU is that images created solely by AI without meaningful human creative input may not be protectable under copyright — meaning you may not own what you generate, and a competitor can use the same image without recourse. This is a live legal area, not settled law.
Platform terms vary significantly
- Midjourney: Commercial use permitted on paid plans. Midjourney retains a license to use your generations. No explicit model-release for AI people.
- DALL-E 3 / OpenAI: You own the output, commercial use permitted. OpenAI may use images for safety training.
- Adobe Firefly: Currently the strongest commercial indemnification. Adobe trains only on licensed content and offers IP indemnification for enterprise customers — making it the default choice for regulated industries or risk-averse brands.
- Stable Diffusion (self-hosted): You control everything. Copyright and IP responsibility sits entirely with you.
Our recommendation for most clients is to default to Adobe Firefly for externally-facing commercial use, and use Midjourney / DALL-E for concept stages where the image won’t be published directly.
Building a prompt discipline that produces consistent brand output
The difference between an agency that gets good results from AI imagery and one that gets frustratingly inconsistent results is almost always prompt discipline. Here’s the structure we use internally:
The four-part prompt framework
- Subject — what is in the image, described precisely
- Environment — setting, lighting quality, time of day
- Technical style — camera, lens, grain, aspect ratio (e.g., “shot on 85mm, slight grain, 4:5 ratio”)
- Brand anchors — colour temperature, mood words that lock the emotional register
We maintain a prompt template document per brand. When a new campaign starts, we spend 30–60 minutes dialling in the template with test generations, then add it to the brand’s content library. Anyone on the team can generate on-brand imagery from that template without needing to rebuild the prompt from scratch each time.
- Consistency across team members increases significantly
- Prompt templates also make it easier to train new team members on brand standards
- You can version-control prompt templates as you would any other brand asset
For brands that want to go deeper on this — aligning AI image generation with a wider brand identity system — the prompt template is part of what we build out in a visual identity engagement. AI readiness is something we now explicitly consider when building brand guidelines.
Our decision framework: when to use AI vs. book a shoot
This is the practical output of our field testing. Use it as a starting filter, not a rigid rule.
Use AI when
- The image is atmospheric, abstract, or supports text without needing to depict reality
- You need volume (20+ unique images) and budget doesn’t allow traditional photography at that scale
- You’re in concept or testing phase — not final production
- The image will be seen at web resolution, not large print
- No specific product, face, or identifiable location needs to be accurate
Commission real photography when
- Your actual product must be in the image accurately
- Real people — team, customers, models — are part of the scene
- You need documentary credibility (case studies, testimonials, events)
- The image will be used at large print scale or in high-scrutiny contexts
- The industry carries regulatory considerations (medical, financial, food safety)
Use AI-enhanced photography (hybrid) when
- You have real product photos but need different backgrounds, seasons, or settings at scale
- You have real portraits but need environmental variation
- You want to extend a limited shoot across more contexts than were originally captured
What this means for AI content strategy more broadly
AI photography is one piece of a larger shift in how marketing content is produced. The teams that get the most out of it are the ones who’ve thought clearly about which parts of their content system benefit from AI acceleration and which parts still need human craft and legal accountability. It’s the same thinking we apply when we help clients build out an AI content creation workflow — the point is never to replace good creative judgment, it’s to put human time where human judgment actually matters.
Imagery is a good test case because the failure modes are so visible. A bad AI image in a campaign is immediately apparent to anyone who looks carefully. That visibility makes it a useful discipline: if your team can make good decisions about when to use AI-generated photography and when not to, they’ll make equally good decisions about the rest of the AI stack.
The short version is this: AI-generated photography is a genuinely powerful tool for the right use cases, and a quiet liability for the wrong ones. The teams winning with it right now aren’t using it everywhere — they’re using it precisely, with a clear-eyed understanding of where it creates value and where it creates risk.
If you want to work through what the right balance looks like for your brand, reach out and tell us where you are. We’re happy to look at your current content operation and give you a straight answer on where AI imagery fits.
