July 7, 2026 · 6 min read

Does the score match your eye?

Our Estimated Preference Score puts a single aesthetic number on every image. It's a faithful summary of the HPSv3 preference model — but a summary of a crowd's average taste is not the same as what you see. So before you trust the number, test it against your own eyes.

Take the blind test

Below are pairs of images generated from the same prompt by two different models, drawn as a fair random sample across five categories. Model names and scores are hidden. Just pick the one you prefer — then we reveal which one the metric scored higher, and keep a running tally of how often you two agree.

Pair 1 / 20 · Spatial Reasoning

Prompt: An astronaut in a spacesuit riding a bicycle through a park

Click the image you prefer.

Why you and the metric drift apart

This isn't the metric being “wrong.” HPSv3 is trained on HPDv3, a large dataset of aggregate human preference — the averaged verdict of many raters. The EPS number faithfully reflects that average (we validated it matches the underlying pairwise preferences on 97.8% of model pairs). But an average is a blunt instrument for a single decision: on any one pair, the “better” image is a matter of taste, genre, and what you happen to be looking for. The score is a useful prior — not a verdict.

It also depends on what you're making

The single number hides another thing: aesthetic strength is category-specific. Scored per category, the median model swings about nine rank positions depending on the category. qwen-image is #1 for spatial-reasoning scenes and #1 for graphic design, but #11 for human realism. seedream-v4 is top-five on people and scenes yet drops to #22 on studio shots. The best model for your use case is often not the one on top of the overall list.

Where the metric is most confident

These are hand-picked pairs with the largest score gaps — the cases where the metric is most sure. The image the metric scored far higher is marked. Do you agree every time?

Human realism — A 5-year-old child with round cheeks and large eyes, crying with tears streaming down their face

Human realism — A person standing with both hands visible at their sides, fingers relaxed and naturally spread

Human realism — A person counting to three on their fingers, with index, middle, and ring fingers extended

Professional Studio — A street photograph with shallow depth of field as if shot at 50mm f/1.8, subject sharp with soft city lights in the background

Spatial Reasoning — A group photo of an elephant, a horse, a dog, a cat, and a mouse, all standing in a line with correct real-world proportional sizes

Spatial Reasoning — A purple carrot, a red tree with red leaves, and a white watermelon sliced open

A systematic tell: it prefers a photo over the style you asked for

The disagreements aren't random — the metric leans hard toward photorealism. Both prompts below explicitly asked for a non-photographic style. In each, the model that correctly rendered what was asked scored far below one that ignored the brief and produced a glossy photo or painting. The metric rewarded the wrong answer.

Asked for pixel art on a 32×32 grid — A golden retriever sitting in a garden, rendered in pixel art style with a 32x32 grid and limited color palette

Asked for flat-colour Japanese anime — A golden retriever sitting in a garden, rendered in Japanese anime style with flat colors and large expressive eyes

It isn't a fluke of these two prompts — a photoreal-leaning model tops most of the style-diversity set. If you're generating stylised art, the aesthetic score is worse than useless here: it actively rewards models for not doing what you asked.

How to use the score

Treat Overall and EPS as a coarse filter — they reliably separate the strong models from the weak — and then look with your own eyes for the job you actually have. Every generated image on ImageBench is published for exactly this reason. Browse them in the gallery and judge for yourself.