How 20 AI image models render: “A classroom with exactly four desks in two rows, a globe on the teacher's desk, a chalkboard with a single drawn triangle, empty of people.”
Same prompt, sent verbatim, at each model's default settings and a fixed seed. Judge: every stated element present, correct count, correct placement. 125 community votes so far.
Core suite prompt · voting closed 2026-09-19
Checklist: exactly four desks · two rows · globe on teacher's desk · chalkboard with a single triangle · no people
20 text-to-image models rendered this prompt adherence prompt with the identical text at their default settings and a fixed seed. GPT Image 2.5 Sunburst is the community favourite so far, winning 71% of its 14 head-to-head votes on this prompt. Grok Imagine Image Quality was the fastest at 4.9s.
What to look for: the right count of objects, the right spatial relationships, the stated colours and attributes, and nothing added.
- GPT Image 2.5 Sunburst71%OpenAI21.9s · 80 cr Try
- Muse Image 1.071%Meta9.7s · 12 cr Try
- Kling v3 Omni62%Kuaishou40.3s · 35 cr Try
- Ideogram v460%Ideogram14.3s · 40 cr Try
- Krea 2 Large57%Krea38.3s · 80 cr Try
- Qwen Image 3.0 Pro54%Alibaba14.2s · 36 cr Try
- GPT Image 253%OpenAI · retired from voting49.7s · 60 cr Try
- ImagineArt 253%ImagineArt · retired from voting32.0s · 60 cr Try
- 4.9s · 80 cr Try
- Recraft v450%Recraft · retired from voting19.8s · 45 cr Try
- ImagineArt 1.5 Pro47%ImagineArt · retired from voting51.2s · 55 cr Try
- Nano Banana 2 Lite44%Google9.9s · 36 cr Try
- 44%16.8s · 50 cr Try
- Krea 2 Medium43%Krea14.8s · 40 cr Try
- GPT Image 2.5 Flare38%OpenAI23.0s · 40 cr Try
- Recraft v4 Pro33%Recraft · retired from voting21.5s · 300 cr Try
- Qwen Image 3.023%Alibaba7.6s · 32 cr Try
- Seedream 5.0 Pro20%ByteDance31.7s · 49 cr Try
- Midjourney Niji 714%Midjourney · retired from voting105.2s · 160 cr Try
FAQ
- Which AI model renders this prompt best?
- GPT Image 2.5 Sunburst leads this prompt in the BudgetPixel AI Image Arena, winning 71% of its 14 blind head-to-head votes. Vote on the arena page to add your judgement.
- Were the prompts adjusted per model?
- No. Every model received the exact same prompt, unmodified, at its default settings, with no negative prompt or style suffix. Some models rewrite prompts internally, and that is part of what is being compared.
- Can I run this prompt myself?
- Yes. Each tile’s “Try” link opens the BudgetPixel workshop with this prompt and that model preselected.