← All guides Workflow

Checking images in bulk: a batch workflow

How to review twenty-five images at once, what to look for in a set rather than a single result, and why the outlier matters more than the highest score.

· 8 min read · Best AI Image Detector

Read the set, not the scores. Images taken in one session on one device cluster together, so the picture that sits outside the cluster is the one worth opening first.

Why sets are more informative than singles

Photographs taken by one person, on one device, in one session, share a processing history. They were compressed the same way, denoised by the same software and exported by the same app, so they land in a similar range.

That gives you a baseline for free. You are no longer asking whether 58 is high, which depends on the camera and the compression. You are asking why one image scored 58 when its eleven neighbours scored between 18 and 26.

The comparison removes most of what makes a single score hard to read. Device processing, platform recompression and lighting affect the whole set together, so they cancel out and what remains is genuinely about that one picture.

Scores across one submitted set
Photos 1 to 4
21
Photos 5 to 8
26
Photos 9 to 11
23
Photo 12
74

Illustrative. The absolute values matter less than the gap between the group and the outlier.

A claim set of twelve photographs from one phone. Eleven cluster; one does not.

The batch workflow

  1. Collect originals, not forwards

    Ask for files from the device rather than images pasted into an email. A set where some images have been recompressed and others have not will scatter for reasons that have nothing to do with manipulation.

  2. Run the whole set at once

    Up to twenty-five per batch. Running them together is what gives you the comparison; running them one at a time throws it away.

  3. Read the summary before opening anything

    The batch view shows the spread and how many landed in each band. That tells you whether you are looking at a clean set, a degraded set, or a set with one problem in it.

  4. Open the outlier first

    Not the highest score. The one furthest from its neighbours, which is not always the same image.

  5. Check the region map on anything you open

    Whether the whole frame is hot or one tile is decides what kind of finding you have and what to ask for next.

  6. Export the batch report

    One PDF summarising every image, and a combined JSON with the full result for each. That is the artefact that goes in the file.

What the shape of a set tells you

Reading a batch summary
Pattern across the setUsually meansWhat to do
All low, tightly groupedA clean set from one deviceNothing further
All middling, tightly groupedHeavy processing or platform recompressionAsk for originals before concluding anything
Mostly low, one clear outlierOne image from a different sourceOpen the outlier and read its region map
Widely scattered with no clusterMixed sources or mixed compressionThe set cannot be compared. Treat each separately
All high, tightly groupedEither wholly generated, or a heavily filtered setCheck one image closely to tell which

The fourth row is worth dwelling on. A scattered set is not a set at all for comparison purposes, and the useful response is to find out why the images differ before reading any of the numbers.

Where the limits are

Twenty-five per batch is a browser memory limit rather than a commercial one. Each image is decoded and scored in the tab, and a larger batch on a modest device runs out of room rather than slowing down gracefully.

For a set of two hundred, split it into groups that make sense: by day, by device, by source. That is better practice anyway, because comparing images from different sessions reintroduces exactly the variation the batch approach removes.

Who this workflow suits

  • Claims handlers. A claim arrives as a set, and the outlier is the standard finding.
  • Moderation teams. Ordering a review queue rather than judging items individually.
  • Agencies. Checking a delivered batch of assets before it goes to a client.
  • Finance teams. A single expense claim with twenty receipts photographed in one session.
  • Researchers. Scoring a sample and looking at the distribution rather than any single result.

Building a baseline for a source you see often

Teams that review images from the same places repeatedly can go further than comparing within one set. Recording where a source normally lands turns every future batch into a comparison against history rather than against itself.

A particular claims portal that recompresses heavily might produce scores clustering around 45. That is not suspicious, it is that portal. Knowing the number means a set arriving at 45 raises nothing and a set arriving at 70 raises a question immediately.

Build it from images you have already resolved rather than from a sample you have not checked. Twenty known-good sets per source is enough to see the range, and the exercise usually surfaces a processing quirk somebody had been misreading as a signal for months.

Keep the note somewhere the whole team can read it. A baseline held in one person’s head stops being a baseline the week they are on leave.

Questions people ask

How many images can I check at once?
Twenty-five per batch. That is a practical ceiling on what a browser tab can decode and score without running short of memory, rather than a limit set to sell an upgrade. Larger sets split naturally by day, device or source, which is better practice anyway.
Should I look at the highest score or the outlier?
The outlier, and they are often different images. A set where everything scores around 60 because the photos were heavily compressed contains no finding. A set clustered at 20 with one image at 55 contains one, even though 55 is the lower number.
Does batching affect accuracy?
No. Each image is scored independently by the same model with the same settings. The batch view adds a comparison across results; it does not change how any individual result is produced.
Can I export a whole batch?
Yes. A summary PDF listing every image with its score and verdict, and a combined JSON containing the full result for each including every tile value. For a claim or a case file, both are worth keeping.
What if the set has mixed file types?
Formats do not matter; processing history does. A batch of HEIC files straight from one phone compares cleanly. A batch mixing camera originals with social media downloads does not, whatever the file extensions say.
How long does a batch take?
Roughly four seconds per image after the model has loaded, so a full batch of twenty-five is a couple of minutes. The first image in the first batch of a session is slower, because the model downloads before anything can run.