We Replaced Our Stock Photo Budget with AI for 30 Days
Updated · May 6, 2026
Our stock photo bill hit $340 last October and nobody on the team could tell us exactly why. We were running a Shutterstock subscription we’d upgraded once and never reviewed, plus occasional Getty Images licenses for shots Shutterstock didn’t carry. We publish around 30 blog posts, 60 social assets, and 12 newsletter headers per month — roughly $8.50 per image we actually used. In March, we canceled both subscriptions and gave ourselves 30 days with AI generators only.
The rules we set for ourselves
No safety net. That was the core rule. No downloading a stock image “just this once” when a prompt wasn’t working. Every piece of visual content for publication had to come from an AI generator or our own existing archive of brand photography.
We committed to logging every image: tool used, number of prompts, and time spent per final asset. If something genuinely couldn’t be generated — a screenshot of real software, an image tied to a specific news event, a photo of an actual product — we’d use a text-based placeholder or skip the image rather than sneak in a stock purchase. Our quality bar stayed the same. We weren’t trying to make AI images look like AI images.
How we picked the tools (and one we almost used)
We almost turned this into a full evaluation process before the experiment even started. There are a dozen generators worth considering, and the temptation to spend two weeks comparing them was real. We didn’t. We picked three tools with clearly differentiated strengths and started the clock.
Midjourney was the anchor. At $30/month for the Standard plan, it consistently produces the most polished output for editorial and concept imagery — abstract technology illustrations, moody header images, flat-lay-style compositions without actual products. The prompt learning curve is real, but we’d already spent time with it before this experiment.
Leonardo AI handled volume. On days when we needed eight or ten images and weren’t precious about prompt iteration, Leonardo’s free tier (150 tokens daily) covered the overflow. We upgraded to their $12/month plan midway through the month when we started hitting limits consistently.
Adobe Firefly came along for free — bundled in our existing Creative Cloud subscription. We used it primarily for images that needed to match brand colors precisely, since Firefly integrates directly into Photoshop and makes targeted adjustments faster than re-prompting from scratch.
We tested DALL-E 3 via ChatGPT for two days and dropped it. The outputs were acceptable but not controllable enough for production use — we couldn’t reliably achieve the same aesthetic twice. Stable Diffusion was on the shortlist, but two hours of local setup work convinced us the output-to-effort ratio wasn’t right for a team workflow where speed matters.
What actually worked? Our results by week three
By day 21 we had a functional rhythm. In our testing, abstract concept images — data visualizations, hands interacting with glowing interfaces, minimal workspace scenes — averaged four Midjourney prompts and about six minutes per final published asset. We were spending comparable time searching Shutterstock for that category of image anyway.
Blog post hero images improved in one measurable way: originality. Stock photo aesthetics are recognizable in the worst sense. The “diverse team celebrating around a laptop” image has appeared in tens of thousands of articles. Our AI-generated equivalents were at least novel.
The workflow that locked in by week three: rough prompt in Midjourney, upscale the best result, import to Canva for text overlays and brand-color adjustments. Canva’s built-in AI image tool handled quick social media variants without leaving the design editor — useful since we were already paying for Canva Pro at $15/month.
One result we didn’t anticipate: building a shared prompt template library in week two made week three significantly faster. Twelve tested prompt structures that consistently produced on-brand results. That document is now a permanent team asset, independent of anything this experiment proved.
Can AI generators handle images of people?
This is where the experiment ran into its clearest wall. Consistently, reliably, at the quality level professional stock photography achieves — no, not yet.
We needed images of professionals in office environments for four articles in week two. Across dozens of prompts and multiple generators, results ranged from plausible-but-slightly-off to full uncanny valley. Hands were wrong. Faces in group shots had subtle inconsistencies that a general audience might not articulate but would definitely feel. Corporate environments looked like concept art for a film about corporations rather than actual workplaces.
Diversity representation was a separate problem. Prompting for “diverse team of professionals” produced outputs that felt algorithmically distributed rather than naturally composed. It’s a genuine limitation, not a minor nitpick. If your audience works in professional environments and looks at them every day, they’ll notice something is off before they can say what it is.
We ended up pulling six images from our 2024 brand shoot archive for articles that required real people. Two articles ran with illustrated alternatives. One article ran without a header image at all, and engagement metrics didn’t noticeably shift. That last data point was more interesting than we expected.
What we’d change next time
Build the prompt library before day one. We wasted roughly four hours in the first week re-discovering prompt structures that worked. That dead time ate into the efficiency gains we were supposed to be capturing, and it would be easy to eliminate in a second run.
We’d also keep a small stock allowance — something like $29/month on Shutterstock’s basic plan — specifically for people photography. Not because AI-generated people look obviously synthetic (they’ve improved significantly), but because the prompting time required to get professional-quality human imagery doesn’t justify the savings when a 30-second stock search produces a usable result.
The framing that matters most: “AI-only” and “stock-only” are both wrong. The useful question is which categories AI handles well enough that paying $249/month for an unlimited subscription license is genuinely wasteful. That list is longer than we expected when we started.
The final stack
After 30 days, our actual monthly spend on visual assets:
- Midjourney Standard — $30/month — hero images, concept illustrations, abstract visuals
- Leonardo AI Pro — $12/month — high-volume days, style consistency work
- Canva Pro (existing subscription) — $15/month — social templates, text overlays, AI image variants
- Adobe Firefly (existing Creative Cloud) — $0 incremental — brand-matched adjustments in Photoshop
- Shutterstock — $0 (canceled)
- Getty Images — $0 (canceled)
Total: $57/month, down from $340. That’s $283/month saved, or roughly $3,400 a year. The offset is approximately three additional hours per week in prompting and iteration work. Whether that trade makes financial sense depends entirely on your team’s hourly rate and content volume. For a solo creator, the Midjourney Standard plan alone at $30/month is probably enough to handle 90% of the use cases we’ve described here.
Frequently asked questions
Did image quality actually suffer compared to stock photography?
For abstract and conceptual imagery, no — and in several cases the output was more original than anything we’d surface on Shutterstock. For images featuring people in realistic professional settings, quality dropped meaningfully and required either more prompting time or alternative approaches like illustrated replacements.
How much prompting experience do you need for this to work?
Enough to understand that specificity matters more than creativity. “A clean minimal illustration of a laptop with abstract data visualizations, blue and white palette, flat design style” consistently outperforms “a cool tech image.” Two or three hours of deliberate practice will get most people to a functional baseline for editorial content.
Are AI-generated images legally safe for commercial use?
Midjourney, Adobe Firefly, and Leonardo AI all include commercial licenses on their paid plans — verify the specific tier your usage requires. Firefly is trained exclusively on licensed Adobe Stock content, making it the lowest-risk option for organizations with legal teams paying close attention to IP exposure.
The experiment confirmed what we suspected but couldn’t quantify: most of what we were paying Shutterstock for was the convenience of fast search, not access to imagery we couldn’t otherwise create. A hybrid approach — AI generators for abstract and conceptual work, a minimal stock allowance for people photography — cuts the bill by more than 80% while preserving quality where it actually matters. The $340/month subscription was a legacy cost, and it took 30 days of friction to finally prove it.
Related reads
- Canva AI vs Adobe Firefly: Which Is Better for Designers?
- We Benchmarked 5 AI Image Upscalers on the Same Photo
- Midjourney vs DALL-E 3 vs Leonardo AI: 2026 Verdict
This article contains affiliate links. If you subscribe through one, we may earn a commission at no extra cost to you. It never changes what we recommend — we only link to tools we actually use. Full disclosure.
