Text-to-image models, measured

Best for each job

• Best overallSeedream 5.0 Pro97 · tied with Nano Banana Pro, GPT-Image 1.5
• Best for peopleGPT-Image 297 · tied with 3 others
• Best for productNano Banana 2.198 · tied with 7 others
• Best for textSeedream 5.0 Flash98 · tied with 6 others
• Best for illustrationGPT-Image 1.598
• Best for sceneGPT-Image 298 · tied with 7 others
• Best for compositionSeedream 5.0 Flash98 · tied with 5 others
• Best valueSeedream 5.096.3 pts/credit
• FastestQwen Image 2.05.9 s

Scoreboard

Ranked by unrounded overall quality, then price; models within 1 point are effectively tied. Category cells are the mean of that category's two prompts. Credits are per image at the base resolution tier; USD is at the Premium plan rate of $0.19 per credit. Higher resolutions can cost more, see pricing.

Same prompt, every model

The images behind the scores, ranked best to worst within each prompt. All of them are AI-generated outputs from the benchmark run, untouched.

PeopleStudio portrait of a woman in her forties with silver-streaked hair, soft window light from the left, neutral grey background, 85mm look, natural skin texture

PeopleA barista in a linen apron laughing while pouring latte art, warm cafe interior, shallow depth of field, candid editorial style

ProductA matte black ceramic coffee mug on a walnut table, single soft light from the left, minimal studio backdrop, product photography

ProductA pair of white running shoes floating mid-air over a pastel pink background, even studio lighting, catalogue shot

TextA vintage cinema marquee at dusk with the words OPEN ALL NIGHT in bold red letters, neon glow, wet street reflections

TextA minimalist poster with the single word BALANCE in large serif type centred on cream paper, one small green circle below the word

IllustrationAnime-style illustration of a young woman with a red umbrella standing on a rainy city street at night, cel shading, vibrant neon reflections

IllustrationFlat vector illustration of a cosy reading nook with a cat asleep on an armchair, limited palette of four colours, clean geometric shapes

SceneA modern glass house on a cliff at golden hour, long shadows, calm ocean below, architectural photography

SceneMisty pine forest at dawn with rays of light through the trees and a narrow wooden boardwalk, wide angle

CompositionThree red apples and one green pear arranged in a single row on a blue plate, top-down view, white tablecloth

CompositionA small orange robot holding a yellow umbrella standing to the left of a tall blue mailbox, overcast day, photoreal

How the benchmark works

• One suite, every model12 fixed prompts, two per category. Square format, base tier, no negative prompt, no seed, no retries. What the model returns first is what gets scored.
• Blind scoringA vision model sees only the prompt and the image, never the model name. Prompt adherence 50%, technical quality 30%, aesthetics 20%. Two passes, averaged.
• Real pipeline timingSeconds are measured from submit to finished file through the same API the studio calls, so queueing and delivery count, exactly as a user experiences them.
• Published evidenceEvery image that produced a score is shown above. If a model returns an error instead of an image, that is recorded as no output, not as a low score.

Questions

Which text-to-image model has the best quality?

In benchmark v1.0 (29 October 2026) Seedream 5.0 Pro, Nano Banana Pro, GPT-Image 1.5 finished within 1 point of each other at the top of the GoCrazy Image Quality Index, which is inside the judge's noise, so they are reported as a tie. Seedream 5.0 Pro has the highest unrounded mean at 97.3.

Which text-to-image model is the fastest?

Qwen Image 2.0 had the lowest median time from submit to finished image at 5.9 seconds across the benchmark prompts.

Which text-to-image model is the cheapest?

Nano Banana 2.1 is among the lowest-priced generators at 1 credit per image, about $0.19 on the Premium plan. 16 of the 26 models cost one credit at the base tier.

Which model gives the most quality per credit?

Seedream 5.0: 96 quality points for 1 credit, the best ratio in the index.

How is the Quality Index measured?

Every generator renders the same 12 prompts at matched settings: square format, base tier, no negative prompt, no seed. Each image is scored blind by a vision model on prompt adherence, technical quality and aesthetics, weighted 50/30/20, in two passes that are averaged, then normalised to 0-100. The 6 categories are people, product, text, illustration, scene, composition.

Where do the speed numbers come from?

From the benchmark run itself: wall-clock seconds from submit to finished image through the same API the AI Image Studio uses, so queueing and file delivery are included. They reflect our infrastructure and the provider queue on the day measured.