9,345 AI-generated images uploaded to Adobe Stock: the full record of reaching 878 downloads and $589/month (~¥95,000) over 9 months

Since the figures here are mostly dollar-denominated, it’s worth a mention upfront: every yen conversion below is quoted exactly as the original article converted it at the exchange rate of that particular month. The author himself notes, “since the rate differs month to month, there’s some margin of error in the yen figures,” so comparisons in yen terms across different months aren’t strictly precise.

The record belongs to a freelancer going by “Junpei.” He had tried stock photography once two years earlier and stopped, then restarted in earnest on June 10, 2025. Because the start date was the 10th, his monthly tallies run from the 10th of one month through the 9th of the next. Nine months in, he had reached 9,345 cumulative uploads, 878 monthly downloads, and $589.59/month (~¥95,830).

Nine months of monthly data

His first revenue (first download) happened on day three after starting. From there, a weekday pattern emerged from the outset: a trickle of sales on weekdays, a dip on weekends.

Narrowing down to one sales channel

In month one, he uploaded to five stock photo sites, but by month two he had judged three of them “not worth the effort” and narrowed down to two: Adobe Stock and PIXTA. Reading further into the record, it’s effectively Adobe Stock alone.

PIXTA operates on a gradually expanding monthly upload cap: 30 in month one, 50 in month two, 100 in month three, 300 in month four, 500 in month five — then it stayed at 500 from month six onward. Downloads there were 0 in June, 1 in July, 3 in August, and 14 in September. The rejection rate was 0%, but by month eight, revenue from PIXTA was judged to be roughly 1/200th of Adobe Stock, and uploads there were cut off. By his account, roughly 9,000 images he wanted to upload but couldn’t were sitting in a backlog.

Most of the images were generated with Midjourney, but from month seven onward he also started using ChatGPT-generated images, noting that “ChatGPT’s quality suddenly got a lot better.” Those, too, were downloading without issue, he reports.

The turning point: month five’s “resubmit with just the tags fixed”

Revenue in months four through six was $408.93 → $409.74 → $431.99, nearly flat. In month five in particular, despite adding 1,063 new uploads, downloads only grew from 537 to 558. An increase of just 21. This was a stagnation phase where uploading more images wasn’t translating into more revenue.

Facing this, he ran an experiment: resubmitting rejected images. Of 900 images he re-uploaded, about 50% were approved. He only touched the title and tags. The images themselves were left completely unchanged. The conclusion he drew from this: “similarity and quality don’t seem to matter”, what determines pass or fail in review isn’t the image content, it’s the metadata.

What did this discovery change? He had over 3,000 rejected images sitting idle, until then, dead inventory. If rewriting metadata got 40–50% of them through, that turned into stock he could grow without generating anything new. In fact, in month eight he “spent the whole time resubmitting rejected images instead of generating new ones,” and that same month, downloads jumped from 539 to 790 and revenue jumped from $401.90 to $616.62 (an increase of roughly ¥33,000). This overlaps with the timing when month seven’s 1,369 new uploads would also be kicking in, so the effect of resubmission alone can’t be cleanly isolated, but it clearly contributed to breaking through the stagnation.

For reference, the rejection rate for newly generated images was about 45%, while the rejection rate for resubmitted images was over 60%. Even so, “nearly 40% of once-rejected images passed review the second time.”

Breaking down the growth

Revenue in this business is determined by per-unit price times download count. On the price side, he has almost no discretion. Adobe Stock’s minimum per-image rate is $0.33, rising to $0.36 once cumulative downloads pass 1,000, and to $0.38 past 10,000. Outside of that, the base rates are $0.66 and $0.99, but that $0.99 tier fluctuates, sometimes dropping to $0.91 in a given month or rising to $1.04 around New Year’s. He attributes month nine’s revenue drop, despite an increase of 88 downloads, to this rate fluctuation combined with zero extended-license sales that month.

Extended licenses first occurred in month four, a single one, worth $26.40, 30 to 80 times the standard per-unit rate. The odds are low, but he notes his expectation that the frequency of hitting one should rise as the number of uploaded images grows.

Download counts also don’t scale proportionally with upload counts. Month five’s “1,063 new uploads for +21 downloads” is the clearest example of this; conversely, in months five and six, he prioritized blog writing and added almost no new images, yet revenue didn’t fall, 603 downloads and $431.99. This is clear confirmation of the stock-type nature of this business: existing inventory keeps selling on its own.

His observations about buying patterns are also concrete: “some people download a batch of images all in the same genre at once”, bulk purchases happen by genre, in categories like weddings or beauty. From this, he’s derived an operating policy: rather than hunting for underserved niches, it’s better to lean into popular genres and expand variations horizontally. That policy aligns with his earlier conclusion that similarity between images doesn’t need to be a concern.

What isn’t going well

Review itself is the bottleneck. As of month three, over 2,300 images were awaiting review, and he estimates the monthly upload ceiling is roughly 1,000. On top of that, 101 images have simply never been reviewed no matter how long he waits, deleting and resubmitting them hasn’t helped either. The reason is unknown.

There’s also a large gap between generated volume and uploaded/accepted volume. In month one, he generated roughly 2,200 images, uploaded 1,050, and only 534 made it into the accepted catalog. By month nine, he had nearly 8,000 unused generated images sitting around, plus over 11,000 reference images he draws inspiration from. The bottleneck in this business is the work of getting images through review and posted, not the work of making them.

Even though he found that a mass tag correction was effective, that correction work itself turned out to be heavy. Adobe Stock lets you bulk-copy-paste tags when uploading new images, but editing tags after the fact has to be done one image at a time. He estimated it would “take about three full months,” and then admitted, the following month, “it got tedious so I gave up on it.”

How much of this is copyable

The replicable parts are clear. Startup cost is limited to subscriptions for an image-generation AI and upscaling software, no inventory, no shipping. Getting the first download on day three shows there’s no need to build an audience, the buyers already exist on the other side of Adobe Stock’s search bar, and the entire step of building your own audience is skipped. The discovery that metadata determines approval and visibility is also directly transferable knowledge.

What’s hard to replicate is the sheer volume of throughput itself. 9,345 accepted uploads over 9 months, with generation volume far exceeding that. Sustaining this workload is premised on his position as a freelancer who can allocate daytime hours as he sees fit, and on a work style where he can shift priorities month to month alongside running a blog. Being able to have a month like five and six, “zero new uploads because I prioritized the blog”, is only possible because this side business isn’t covering his living expenses.

Platform dependence is also unavoidable. Over 99% of revenue is concentrated in a single company, which also sets both the review standards and the price table. Even he can’t tell why the per-unit rate shifted from $0.99 to $0.91. This business, having reached roughly ¥95,000/month in nine months, is simultaneously a business with no control over its own pricing.

Related reading

• Photo AI — a leading example of an individual developer turning AI-generated images into a product. The contrast in sales approach is instructive.
• 24 Kindle books — a record of how volume-stacking content sales translated into revenue over time.