Subscore

QualityTesting methodology

Quality measures how good the finished images actually look.

Most image generators can create one impressive result when everything goes right. We generate a full batch because we want to know whether the app can produce good images consistently—not only one lucky image for its homepage.

We check realism, major visual problems, composition, and the highest resolution you can download.

4 evidence groups
4 scored tests

Quality is organized into 4 evidence groups. Each group contains one scored test: Realism, Visual Errors, Composition, and Resolution. All four tests count equally toward the final Quality score.

How the Quality score is calculated

Every test gets a score from 0 to 10.

All four tests count equally. We multiply each test score by 25% and add the points together. Quality makes up 34% of the Images score.

25%Realism
25%Visual Errors
25%Composition
25%Resolution

Weighted evidence groups (combined 100%)Quality score34% of Images score

View exact calculation

Each test gets a score from 0 to 10. We multiply every score by 25% and add the points together.

Quality makes up 34% of the Images score. Images makes up 15% of the overall performance score.

Scored tests and weights

Scored test	How much it counts	See scoring
Realism	25.00%	View
Visual Errors	25.00%	View
Composition	25.00%	View
Resolution	25.00%	View
Total	100%

How the score is calculated

Each test gets a score from 0–10
Test score × how much it counts
Example: 8.40 × 25.00% = 2.10 points
We do this for every test
We add all the points together
Final Quality score
Counts for 34% of Images

Example calculation

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

Scored test	Test score	How much it counts	Calculation	Points added
Realism	8.40	25.00%	8.40 × 25.00%	2.10
Visual Errors	8.00	25.00%	8.00 × 25.00%	2.00
Composition	8.20	25.00%	8.20 × 25.00%	2.05
Resolution	8.00	25.00%	8.00 × 25.00%	2.00
Final Quality score	—	100%	Add all points	8.15/10

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 Quality matters

Images are one of the biggest reasons people sign up for AI girlfriend apps.

The problem is that almost every modern image generator can create something that looks good at first glance. That does not mean it performs well every time.

You may get one great image followed by several with broken hands, damaged faces, strange bodies, or terrible framing.

This becomes even more frustrating when every generation costs credits. A broken image is not only disappointing—it can also mean you need to pay again and hope the next attempt works.

That is why we test a full batch. A strong generator should create good-looking, usable images consistently instead of relying on one lucky result.

What good Quality looks like

A high Quality score means the images usually look polished, have few major visual problems, are framed properly, and can be downloaded at a useful resolution. A lower score means too many images look broken, badly cropped, poorly arranged, or too small to use properly.

Main benchmarks

GoodTypicalWeak

Realism

80%+60–79%Under 60%

Images with major errors

20% or less21–40%Over 40%

Composition

80%+60–79%Under 60%

Minimum useful standard

Realism: 80%+
Visual Errors: 20% or less
Composition: 80%+
Resolution: 1080p or higher

Common red flags

Broken hands, faces, or limbs appear regularly
Subjects are accidentally cropped out of the image
The background looks damaged or objects merge together
One good image hides a batch of weak results
The highest-quality download is still very small
Images look unfinished even when the generation technically succeeds

How we test Quality

We use a paid account and generate 10 test images.

We upload every image to the same worksheet and review them one by one.

For each image, we rate visual realism and overall quality, composition and framing, and major defects or visual problems.

We also download the highest-quality image available and record its exact width and height.

Using a batch of 10 images helps reduce luck. One unusually good or bad result cannot control the whole score.

Limitations and edge cases

Image generation always includes some randomness.

The same prompt can create a strong image once and a much weaker result the next time. Testing 10 images reduces this problem, but it cannot remove randomness completely.

Results can also depend on the type of image. A generator may perform well on close-up portraits but struggle with full-body poses, several people, or complicated backgrounds.

Quality only looks at how the finished images appear. Whether the generator followed the prompt and kept the same character is tested separately under Accuracy.

Image models also change quickly. Our results show how the generator performed on the recorded test date.

Evidence groups

Quality has 4 evidence groups made up of 4 scored tests.

Realism measures how believable and polished the generated images look.

We look at the full image rather than judging only the character’s face.

Realism

25%of Realism (25% total)

How believable and polished the generated images look.

How we test

We rate every image from 1 to 5. When giving the rating, we look at face, body, hands, lighting, and background. A higher rating means the image looks more realistic and has fewer obvious problems.

What counts

A believable face
Natural-looking body proportions
Hands that look complete
Lighting that fits the scene
A background that looks clear and believable
An overall result that feels finished

What does not count

Whether the image followed every part of the prompt
Whether the same character is preserved across several images
Small style choices that are clearly intentional
Personal preference for one art style over another

Result shown

Realism result: 84%

Average rating: 4.2 out of 5

Scoring

A higher average rating means a higher score.

Average rating	Score
1 out of 5	2/10
2 out of 5	4/10
3 out of 5	6/10
4 out of 5	8/10
5 out of 5	10/10

An average rating of 4.2 out of 5 scores 8.4/10.

The exact average is used between these points.

Evidence group 2 of 4

Visual Errors measures how many generated images contain at least one major problem.

A generation counts as having an error when the problem is serious enough to make the result look clearly broken or difficult to use.

Visual Errors

25%of Visual Errors (25% total)

How many generated images contain at least one major problem.

How we test

We inspect all 10 images and complete a defect checklist for each one. An image is marked as having a major error when it includes at least one serious problem.

What counts

Extra or missing limbs
Broken hands
A damaged face
Objects merged together
Broken or disappearing clothing
A badly distorted background
A body that looks clearly deformed

What does not count

A small detail that most people would not notice
A style choice that looks intentional
A missing prompt detail
A weak pose that does not make the image look broken
A generation failure that produced no image, which is tested under Experience

Result shown

Visual error rate: 20%

2 of 10 images had a major error

Scoring

Fewer images with major errors means a higher score.

Images with major errors	Score
0%	10/10
20%	8/10
40%	6/10
60%	4/10
80%	2/10
100%	0/10

A visual error rate of 20% scores 8/10.

The exact percentage is used between these points.

Evidence group 3 of 4

Composition measures how well the subject and background are arranged inside the image.

Even a realistic image can be difficult to use when the character is cut off, placed awkwardly, or surrounded by a messy background.

Composition

25%of Composition (25% total)

How well the subject and background are arranged inside the image.

How we test

We rate every image from 1 to 5. When giving the rating, we look at whether the requested subject is fully visible, accidental cropping, subject placement, background clarity, and overall balance.

What counts

The character is visible as requested
Important body parts are not accidentally cropped
The subject is placed naturally
The background is clear
The full image feels balanced

What does not count

Whether the prompt details are correct
Small creative framing choices
Personal preference for close-up or full-body images
Visual defects already counted under Visual Errors

Result shown

Composition result: 82%

Average rating: 4.1 out of 5

Scoring

A higher average rating means a higher score.

Average rating	Score
1 out of 5	2/10
2 out of 5	4/10
3 out of 5	6/10
4 out of 5	8/10
5 out of 5	10/10

An average rating of 4.1 out of 5 scores 8.2/10.

The exact average is used between these points.

Evidence group 4 of 4

Resolution measures the maximum image size you can download.

A larger image gives you more detail and is easier to crop, edit, or use on a larger screen.

Resolution

25%of Resolution (25% total)

The maximum image size you can download.

How we test

We generate an image using the highest quality setting available. We download the finished file and record its exact width and height in pixels. We use the downloaded file rather than trusting the resolution shown on the pricing or marketing page.

What counts

The highest-quality image normal users can download
The actual width and height of the finished file
Upscaling that is included and available through normal use

What does not count

Resolution promised only in marketing
A preview image that cannot be downloaded
A larger display size that does not change the real file
Third-party upscaling outside the app

Result shown

Maximum resolution: 1920 × 1080

Resolution level: 1080p

Scoring

We use the ranges below.

Maximum resolution	Score
480p	4/10
720p	6/10
1080p	8/10
4K	10/10

A maximum resolution of 1080p scores 8/10.

Back to Images