How we test AI image models

What stays the same in our image-model tests, how we count outcomes, and what the results can tell you.

Last updated: October 9, 2026

The same photo and prompt

In a controlled comparison, each model in a row receives the same saved input photo and the same prompt. We record the image shape, model versions, settings, and test date. Settings that differ between models are stated as limits of the comparison.

Each article explains its question and how many cases it tested. The original input stays separate from generated results. Older image galleries without matching input and prompt records remain historical archives; they cannot establish a model winner.

Count every tested outcome

  • Completed: the model returned an image. This does not mean the image was good.
  • Blocked: a safety or policy response prevented an image.
  • Failed: another recorded error prevented a usable image. The article states if failures are grouped with blocks.
  • Pending or not tested: excluded from the completed test count unless the article says otherwise.

Completion rates divide completed attempts by tested attempts. We keep blocked and failed results visible and do not guess the reason for a safety block when the provider does not explain it.

Read results with their limits

Compare the face, pose, clothes, and scene in the actual outputs. A small portrait test cannot tell you which model is best for every task. Even the same photo and prompt can produce a different result on a later run, especially after a model or safety-system update.

Prompt collections are selected examples, not controlled model tests. Each guide should explain why its examples fit the task and link to their source packs. Inclusion is not a quality guarantee or safety certification. The method and result labels in a specific article take precedence over this general guide.

Technical records for controlled tests

Each row records the exact prompt and its SHA-256, input storage key and SHA-256, profile ID, aspect ratio, comparison run ID, and exact model IDs. Every successful or blocked attempt must repeat those values. Requests with missing or mismatched proof are stopped before submission.

Article records preserve outputs, outcomes, settings, exclusions, and sample sizes. A rerun keeps the original input and prompt records and gets a new run ID. Missing proof in an old archive is not filled in by assuming the inputs matched.

See a recorded comparison

These articles show the July 2026 controlled rerun, with its own inputs, outcomes, and limits.