Taste Bench · Model Evals · Sept 2026
Any AI can make an image.Taste is what makes an ad.
8 leading image models, each one running inside Omneky. 59 real ad briefs across 4 brands and 5 languages. 1,371 blind reviews, scoring what an ad lives or dies on: your logo, your product, your words, your market.
Last updated: Re-run as new models ship. The harness stays on for every model.
The leaderboard
Average quality score out of 10 across all judges and briefs.
- 17.4
GPT Image 2.5 FlareOpenAI - 27.4GPT Image 2.5 SunburstOpenAI
- 37.4GPT Image 2OpenAI
- 47.2
Nano Banana ProGoogle - 57.1
Grok Imagine 2.0xAI - 67.0Nano Banana 2Google
- 77.0
Muse ImageMeta - 86.4
Seedream 5 ProByteDance
- 8
- Image models tested
- 469
- Ads generated
- 1,371
- Blind judge reviews
- 5
- Languages
These scores are Omneky, not the model's own app.
Every model on this page ran inside Omneky's image harness. Omneky read the brand, wrote the art direction, and handed the model your logo, your product, and the brief. Gemini doing well here means Gemini draws well after that work. Opening the Gemini app does not give you this.
Leaderboard
The leaderboard
Same Omneky harness for every row: same brand kit, same brief, one shot. Scored blind. This is not the Gemini app, and it is not ChatGPT.
59 briefs in this view. Click a row for the full breakdown.
- Ad score
- 7.4
- Ready to run
- 42%
- Ad score
- 7.4
- Ready to run
- 54%
- Ad score
- 7.4
- Ready to run
- 37%
- Ad score
- 7.2
- Ready to run
- 32%
- Ad score
- 7.1
- Ready to run
- 30%
- Ad score
- 7.0
- Ready to run
- 32%
- Ad score
- 7.0
- Ready to run
- 34%
- Ad score
- 6.4
- Ready to run
- 22%
First try, with Omneky's review step switched off so the models can be compared. The harness stayed on: Omneky still read the brand and wrote the art direction for every model, including Gemini.
Grok Imagine 2.0 couldn't run 3 briefs: they included more input images than it accepts.
Every brief
However you start, you get an ad.
Each row is one brief. Every model in that row got it from Omneky, with the same logo, product, and art direction. A chat app does not start from here.
| Brief | GPT Image 2.5 FlareOpenAI | GPT Image 2.5 SunburstOpenAI | GPT Image 2OpenAI | Nano Banana ProGoogle | Grok Imagine 2.0xAI | Nano Banana 2Google | Muse ImageMeta | Seedream 5 ProByteDance |
|---|---|---|---|---|---|---|---|---|
No prompt Asarai Arabic · Format 1:1 Nothing. Omneky started from the brand alone. | ||||||||
No prompt Asarai English · Format 1:1 Nothing. Omneky started from the brand alone. | ||||||||
No prompt Asarai Spanish · Format 1:1 Nothing. Omneky started from the brand alone. | ||||||||
No prompt Asarai Hindi · Format 1:1 Nothing. Omneky started from the brand alone. | ||||||||
No prompt Boco to Deco Japanese · Format 1:1 Nothing. Omneky started from the brand alone. |
Click any ad to see how each judge scored it.
The checks
A pretty picture isn't an ad.
Most image benchmarks score beauty. The Taste Bench scores taste: whether you could run the ad tomorrow.
Your logo, not a lookalike
Copied exactly. Not redrawn, recolored or swapped for something close.
Your product, exactly as it ships
The real pack, label and shape, not an AI guess at what it might look like.
Your headline, spelled right
Word for word when you give us copy. Short and clean when you don't.
Your language and script
Arabic reads right to left. Hindi is in Devanagari. Japanese stays Japanese.
No invented discounts, stats or reviews
A made-up "50% off" or a fake five-star quote is a compliance problem, not a creative choice.
Clear of the platform's buttons and captions
The logo, the copy and the product stay where Meta and TikTok won't cover them.
Miss any one and it's not an ad. It's a liability.
We don't just test for these. Omneky checks for all of them automatically, so you don't have to.
No model wins every brief.
The model that nails a detailed brief stumbles on a blank one. The best at English copy isn't the best in Arabic. Keeping up is a full-time research job. It's ours, not yours.
Best by starting point
GPT Image 2.5 FlareOpenAIGPT Image 2OpenAIGPT Image 2.5 FlareOpenAIGPT Image 2OpenAIGPT Image 2.5 SunburstOpenAIGPT Image 2OpenAIGPT Image 2.5 SunburstOpenAIGPT Image 2.5 SunburstOpenAIBest by language
GPT Image 2.5 SunburstOpenAIGPT Image 2.5 FlareOpenAIGPT Image 2.5 SunburstOpenAIGPT Image 2OpenAIThat's why Omneky has no model dropdown. We make the call, and we re-make it every time a new model ships.
Want to see your brand in this grid? Start your 7-day free trial.
Free TrialMethod
How we judge an ad
Judges never learn which model made the ad. Every image is renamed and stripped of anything that gives it away.
Three frontier AI models from three different labs, so no model grades its own homework.
Hard pass/fail checks come first. Then quality scores from 1 (broken) to 10 (agency-grade).
An ad is ready to run only if it passes every check and most judges would run it as-is. A split vote counts as a fail.
The harness is what you get
Every score on this page is a model running inside Omneky. We read your brand and turn it into art direction. The model draws. We publish the ads and the verdicts. We don't publish that recipe. A strong Gemini score is Gemini after the harness, not Gemini out of the box.
Always testing
When a new image model ships, it runs the Taste Bench before it gets anywhere near your ads. This page always shows the latest edition.
Model names and logos are trademarks of their respective owners and are used here only to identify the models tested. Brands shown appear in Omneky's test set.