reve 2.0 ranks #2 but can't match nano banana or mai on real prompts

reve 2.0 vs gpt image 2 vs nano banana 2 vs mai image 2.5

reve 2.0 just landed #2 on the text-to-image leaderboard on artificial analysis. we ran our own test to see how this model holds up

same 7 prompts across all four models – paintings in different styles:

• impressionist – rainy paris boulevard at dusk
• surrealist – a grandfather clock melting into moths
• ukiyo-e – a crane over a snow-covered pine
• abstract expressionist – a storm of ochre, cobalt and red
• renaissance chiaroscuro – an astronomer with a glowing globe
• art nouveau – a woman with hair of vines and irises
• fauvist – a mediterranean harbor in impossible colors

cost (7 images):
1. reve – $0.22
2. mai – $0.44
3. gpt – $0.48
4. nano – $0.61

speed (7 images):
1. reve – 57s
2. nano – 1m41s
3. mai – 2m35s
4. gpt – 6m53s

reve wins on price and speed by a wide margin. but the output is a step behind – nice-looking, yet too simple for a #2 model. the crane and the art nouveau maiden are pretty but underdetailed, and the clock and the astronomer come out sloppy

@NanoBanana 2 and @MicrosoftAI image 2.5 handle the same prompts with more depth

so a model ranks #2 and still feels basic the moment you actually use it. that's the problem with benchmarks and leaderboards – the score says nothing about what a model really produces. run the prompts yourself, don't trust the ranking