AI Benchmarks Explained

The largest image the model generates directly, before any separate upscaling step, expressed in megapixels so different aspect ratios can be compared. What you can take from it: for anything going to print or to a large screen, native resolution is the ceiling on quality — upscaling afterwards invents detail rather than recovering it.

For web use almost every current model is already past what you need, so this should not drive the decision.

How reliably the model draws legible, correctly spelled words when a prompt asks for them — a poster headline, a label on a product, a sign in a scene. It is scored separately because it is the single most common failure in image generation and the models differ enormously on it. What you can take from it: if your use is marketing assets, mockups, thumbnails or anything carrying a brand name, this matters more than overall image quality, because a beautiful image with mangled lettering is unusable.

If you only ever generate textless imagery, ignore it entirely.

Whether the image contains what you asked for — the right number of objects, the right colours on the right things, the spatial relationships you specified — rather than something merely in the right style. What you can take from it: this is what decides how many attempts a usable result takes, so it drives real cost far more than the per-image price does.

A model that scores well here can be directed; one that scores badly has to be coaxed, and you pay for every discarded attempt.

People are shown two generated images from the same prompt and pick the better one; the votes become an ELO rating. This is the image-generation category of Design Arena specifically, rather than a model's average across every category it competes in. What you can take from it: it is the only broad public measure of whether an image model produces results people actually prefer, which is not the same as resolution or speed.

Read it beside generation time and cost per image — a model that wins on preference but takes three minutes per image is a different proposition from one that is nearly as good in twenty seconds.

How much video the model produces in a single generation, in seconds. Almost every current model is capped well below the length of a finished piece. What you can take from it: this sets how much stitching your workflow needs. A model capped at five seconds cannot produce a twenty-second shot — you generate four clips and join them, and each join is a place where lighting, faces and motion can visibly jump.

A longer cap is worth real money on any narrative work, and almost nothing on short social cuts.

The tallest frame the model outputs natively — 720, 1080, or 2160 for 4K — before any upscaling. What you can take from it: 1080p is the practical floor for anything client-facing or going to a large screen; 720p is fine for social and for previewing. As with images, upscaling afterwards adds pixels rather than detail, so native resolution is the real ceiling.

Whether the model produces audio along with the picture — dialogue, effects and ambience timed to what is on screen — or returns silent video you have to score separately. Scored 100 for yes, 0 for no. What you can take from it: native audio removes a whole production step and, more importantly, removes the lip-sync problem, which is very hard to fix after the fact.

If your output is silent B-roll it is irrelevant; if anyone on screen speaks, it is close to decisive.