search

Subscore

AccuracyTesting methodology

Accuracy measures whether the image generator creates what you actually asked for.

A picture can look amazing and still be a bad result. You might ask for a red dress and get a black one, or generate the same AI girlfriend again and suddenly receive a completely different face.

We check whether the generator follows your prompt, keeps the character looking the same, and edits only the parts you asked it to change.

  • 6 evidence groups
  • 6 scored tests

Accuracy is organized into 6 evidence groups. Each group contains one scored test: Prompt Accuracy, Character Consistency, Face Consistency, Body Consistency, Style Consistency, and Editing Accuracy.

How the Accuracy score is calculated

Every test gets a score from 0 to 10.

We multiply each test score by how much it counts. We then add all the points together to calculate the final Accuracy score.

The live methodology gives slightly more weight to Prompt Accuracy and the main character-consistency tests. Accuracy makes up of the Images score.

  • 17%Prompt Accuracy
  • 17%Character Consistency
  • 17%Face Consistency
  • 17%Body Consistency
  • 16%Style Consistency
  • 16%Editing Accuracy
Weighted evidence groups (combined 100%)Accuracy score of Images score
View exact calculation

Each test gets a score from 0 to 10. We multiply every score by how much it counts and add the points together.

Accuracy makes up of the Images score. Images makes up 15% of the overall performance score.

Scored tests and weights

Scored testHow much it countsSee scoring
Prompt Accuracy17.00%View
Character Consistency17.00%View
Face Consistency17.00%View
Body Consistency17.00%View
Style Consistency16.00%View
Editing Accuracy16.00%View
Total100%

How the score is calculated

  1. Each test gets a score from 0–10

  2. Test score × how much it counts

    Example: 8.20 × 17.00% = 1.39 points

  3. We do this for every test

  4. We add all the points together

  5. Final Accuracy score

    Counts for 33% of Images

Example calculation

We multiply each test score by how much it counts. We then add all the points together.

Scored testTest scoreHow much it countsCalculationPoints added
Prompt Accuracy8.2017.00%8.20 × 17.00%1.39
Character Consistency8.0017.00%8.00 × 17.00%1.36
Face Consistency8.0017.00%8.00 × 17.00%1.36
Body Consistency6.0017.00%6.00 × 17.00%1.02
Style Consistency8.0016.00%8.00 × 16.00%1.28
Editing Accuracy7.6016.00%7.60 × 16.00%1.22
Final Accuracy score—100%Add all points7.63/10

Special cases

Not Applicable

If a test does not apply, we remove it and spread its weight across the remaining tests.

Unknown

If we cannot verify a result, the test receives a score of 0.

Manual adjustment

In rare cases, we may adjust a score when the calculated result is clearly unfair. We always record the reason.

Why Accuracy matters

Good image quality is not enough when the generator ignores what you asked for.

You may request a red dress, a beach background, and a specific pose. The result might look beautiful but give you a white dress, an indoor setting, and a completely different pose.

This becomes especially frustrating when every generation costs credits. Each wrong result means you may need to pay again and hope the next attempt follows the prompt.

Character consistency matters for the same reason. Your AI girlfriend should not get a different face, body, or age every time you generate another image.

Editing should also be accurate. If you ask to change the outfit, the app should not change the face, body, pose, and background at the same time.

A strong image generator should create what you asked for and keep the character recognizable without making you waste credits on constant retries.

How we test Accuracy

We use a paid account and complete three image tests.

First, we generate a batch of 10 images and rate how well each result follows the prompt.

We then upload a reference portrait and create five variations of the same character. We compare every variation with the reference and check face, body, visual style, and overall character consistency.

Finally, we complete 10 image-editing tasks. Every edit is checked to see whether the requested change was made without changing the face, body, pose, or background.

The active worksheet uses 10 batch images, five character-consistency variations, and 10 editing tasks.

Limitations and edge cases

Image generation always includes some randomness.

The same prompt can create a very accurate result once and ignore several details the next time. Testing several images helps reduce luck, but it cannot remove randomness completely.

Some prompts are also harder than others. A simple portrait is easier to follow than a complicated scene with several characters, objects, poses, and clothing details.

Character consistency can depend on the tools offered by the app. A platform with a proper reference-image feature may perform better than one that only lets you describe the character with text.

A character can also keep the same face but change her body or visual style. That is why we score Face, Body, and Style Consistency separately.

This score measures whether the results are accurate. How attractive, realistic, or visually polished the images look is tested separately under Quality.

Evidence groups

Accuracy has 6 evidence groups made up of 6 scored tests.

17%

Prompt Accuracy

1 scored test

Prompt Accuracy measures how well the generated images follow your instructions.

A beautiful image is not accurate when it leaves out the important details you asked for.

Prompt Accuracy

How well the generated images follow your instructions.

How we test

We generate 10 test images. Each prompt contains five clear details to check, such as clothing, color, location, pose, or an object or background detail. We check how many requested details appear correctly in the finished images.

What counts
  • The requested clothing appears
  • Important colors are correct
  • The location matches the prompt
  • The requested pose is followed
  • Important objects or background details are included
What does not count
  • Extra small details that do not change the request
  • Personal preference about how attractive the result looks
  • Visual quality problems tested under Quality
  • A detail that is only partly visible when the prompt asked for it clearly
Result shown

Prompt Accuracy result: 82%

41 of 50 requested details were followed

Scoring

A higher percentage means a higher score.

Prompt details followedScore
20%2/10
40%4/10
60%6/10
80%8/10
100%10/10

A result of 82% scores 8.2/10.

The exact percentage becomes the score.

Evidence group 2 of 6

17%

Character Consistency

1 scored test

Character Consistency measures whether the character still looks like the same person across different images.

This is the overall consistency result. Face, Body, and Style Consistency show the separate parts behind it.

Character Consistency

Whether the character still looks like the same person across different images.

How we test

We upload one reference portrait. We then generate five new images of the same character and compare every result with the reference. The overall result is calculated from the Face, Body, and Style ratings in the consistency worksheet.

What counts
  • The character is still clearly recognizable
  • The face, body, and style remain similar
  • Important identifying features stay visible
  • The result looks like another image of the same character
What does not count
  • A completely different person with similar hair
  • A face match while the body changes heavily
  • The same art style with a different identity
  • Small pose, clothing, or background changes that were requested
Result shown

Overall consistency result: 80%

Scoring

A higher percentage means a higher score.

Overall consistencyScore
20%2/10
40%4/10
60%6/10
80%8/10
100%10/10

A result of 80% scores 8/10.

The exact percentage becomes the score.

Evidence group 3 of 6

17%

Face Consistency

1 scored test

Face Consistency measures whether the character’s face stays recognizable.

The hairstyle or outfit may change, but the character should not suddenly look like a different person.

Face Consistency

Whether the character’s face stays recognizable.

How we test

We compare the face in each of the five variation images with the reference portrait. Each image is rated based on whether the face matches the reference.

What counts
  • The same main facial structure
  • Similar eyes, nose, mouth, and jaw
  • The character remains easy to recognize
  • Normal changes caused by expression or camera angle
What does not count
  • A different person with similar hair
  • A face that only matches because the pose is copied
  • Large age changes
  • Major changes to facial shape
  • Small lighting changes that do not affect identity
Result shown

Face Consistency result: 80%

4 of 5 images kept the same face

Scoring

A higher match rate means a higher score.

Images with the same faceScore
0 of 50/10
1 of 52/10
2 of 54/10
3 of 56/10
4 of 58/10
5 of 510/10

A result of 4 matching images scores 8/10.

Evidence group 4 of 6

17%

Body Consistency

1 scored test

Body Consistency measures whether the character keeps the same general body type and proportions.

A character should not change from slim to muscular or from short to very tall unless the prompt asks for it.

Body Consistency

Whether the character keeps the same general body type and proportions.

How we test

We compare the body in each of the five variation images with the reference. We check whether the main body features stay consistent.

What counts
  • Similar height appearance
  • Similar body type
  • Similar proportions
  • The same general build
  • Natural changes caused by pose or clothing
What does not count
  • A completely different body type
  • Large changes in height or proportions
  • Unrequested changes to body shape
  • A difference caused only by loose clothing or camera angle
  • Breast-size controls tested separately under Customization
Result shown

Body Consistency result: 60%

3 of 5 images kept the same body

Scoring

A higher match rate means a higher score.

Images with the same bodyScore
0 of 50/10
1 of 52/10
2 of 54/10
3 of 56/10
4 of 58/10
5 of 510/10

A result of 3 matching images scores 6/10.

Evidence group 5 of 6

16%

Style Consistency

1 scored test

Style Consistency measures whether the images keep the requested visual style.

A realistic image should not suddenly look like anime, a 3D render, or a cartoon unless the prompt asks for that change.

Style Consistency

Whether the images keep the requested visual style.

How we test

We compare the visual style of each variation with the reference and the requested style. The active worksheet rates five variation images.

What counts
  • Realistic images remain realistic
  • Anime images keep an anime look
  • Lighting and rendering remain reasonably similar
  • The requested art style is easy to recognize
What does not count
  • Normal changes in lighting or background
  • A different pose
  • Small changes in sharpness
  • A major switch to another art style
  • Style changes clearly requested in the prompt
Result shown

Style Consistency result: 80%

4 of 5 images kept the requested style

Scoring

A higher match rate means a higher score.

Images with the same styleScore
0 of 50/10
1 of 52/10
2 of 54/10
3 of 56/10
4 of 58/10
5 of 510/10

A result of 4 matching images scores 8/10.

Evidence group 6 of 6

16%

Editing Accuracy

1 scored test

Editing Accuracy measures whether the app makes the requested change without damaging the rest of the image.

If you ask to change the outfit, the face, body, pose, and background should stay the same.

Editing Accuracy

Whether the app makes the requested change without damaging the rest of the image.

How we test

We complete 10 image-editing tasks. For every edit, we check whether the requested change was made and whether the face, body, pose, and background stayed the same. This creates 50 checks in total.

What counts
  • The requested change is clearly visible
  • Unrelated parts remain unchanged
  • The character stays recognizable
  • The edited image still looks usable
What does not count
  • The requested change is ignored
  • The whole image is regenerated
  • The face or body changes unnecessarily
  • The pose changes when it was meant to stay the same
  • The background changes without being requested
Result shown

Editing Accuracy result: 76%

38 of 50 editing checks passed

Scoring

A higher percentage means a higher score.

Editing checks passedScore
20%2/10
40%4/10
60%6/10
80%8/10
100%10/10

A result of 76% scores 7.6/10.

The exact percentage becomes the score.

Back to Images