What 3,576 Similar Content Refusals Reveal About Large Uploads and Repeated Setups

The refusals clustered in large uploads

Of the 3,576 refused files, 3,004—84.0%—came from uploads in which at least 20 files received the same Similar Content refusal. This figure shows where the refusals were concentrated. Repetition inside those uploads requires a separate comparison of what the images contained.

For that second comparison, we grouped files by main subject, type of scene, framing, and background. That comparison found repeated setups even when the prompts, framing, backgrounds, titles, or keywords were different.

The sample covers uploads from June 28, 2025 through July 15, 2026. Files without a matching prompt were left out of the prompt-based comparisons. The grouping is our way of comparing this sample, not Adobe's own similarity score.

Different prompts still led to familiar setups

We found 2,263 combinations of subject, scene, framing, and background. Of those, 566 appeared more than once and covered 1,879 files—52.5% of all refusals.

The largest repeated pattern was a wide vineyard scene: 34 refused images came from 25 prompts across 24 uploads. Thirty wooden-table still lifes came from 28 prompts across 24 uploads. Fourteen conventionally framed living rooms each came from a different prompt and upload.

Most refused files did not share an identical prompt. Of the 3,209 distinct prompts linked to refused files, 2,898 (90.3%) were linked to only one refused file. Many different prompts still produced the repeated setups above: familiar layouts that could fill the same kind of article, ad, or design.

The six images below are separate refused files from the sample. They show its range of subjects; they are not edits of one image or a record of what Adobe compared.

Many refused sets already included framing or background changes

We found 209 sets in which at least two refused files from the same upload showed the same main subject. Together, those sets contained 502 files.

Within the 209 sets, 118 (56.5%) used more than one framing or composition, 100 (47.8%) used more than one background, and 164 (78.5%) used at least two different combinations of subject, scene, framing, and background.

This comparison contains only refused files, so it cannot measure whether a particular change improves acceptance. It shows that differences in framing and background were already present within many of the refused sets.

Keyword changes could not change the picture

Among the 3,576 refusals, 678 files (19.0%) shared their full prompt with another refused file, and 666 files (18.6%) shared the same title. Only 55 files (1.5%) repeated an entire keyword list.

Most keyword lists were different. Reordering terms, adding synonyms, or replacing descriptions changes the listing, not the picture itself.

Another 581 files (16.2%) began their keywords with words such as “standard,” “directional,” “diffused,” “shallow,” “degree,” or “topdown”—terms that look more like generation or camera instructions than useful search phrases. Replace them with words that describe the image, but do not expect keyword edits to make a visually repetitive image distinct.

Review the group before submitting more files

The two findings lead to two separate review questions. Large uploads deserve an early check because most refusals in this sample came from them. Within any upload, compare what each image could illustrate and remove files that repeat the same subject and setup.

Do this before submission. If two images could fill the same article, ad, or design with little difference, keep the stronger one instead of relying on a new crop, background, title, or keyword list.