TASTE BENCH:Sunburst 廣告質量 7.38,首稿可投放僅 54.2%
Omneky 用同一套 59 條 brief×四品牌×五語言盲測八模型:GPT Image 2.5 Sunburst 均分最高 7.38/10,但首稿可直接投放只有 54.2%。選型別隻看分——廣告可跑通纔是終審。
示範影片
原帖未提供效果圖
提示詞
Today we're launching TASTE BENCH, Omneky's eval suite for the creative quality of AI-generated ads. For image ads, the top model, GPT Image 2.5 Sunburst, averaged 7.38 out of 10 on quality, yet only 54.2% of its ads were ready to run as-is. Across eight models, ready-to-run rates ranged from 22.0% to 54.2%. Every model got the same 59 briefs across four brands and five languages, and a blind panel of AI judges from OpenAI, Anthropic, and Google checked each ad for exact copy, faithful logos and products, safe zones, and fabricated claims. These are first-attempt numbers with our own review step turned off. We built TASTE BENCH because once you can measure taste, you can hill climb it. It covers image and video models, and you can explore every image ad, brief, judge explanation, and score at http://www.omneky.com/tastebench.
中文翻譯
今天我們正式發佈 TASTE BENCH,這是 Omneky 針對 AI 生成廣告創意質量的評估套件。在圖片廣告方面,表現最好的模型 GPT Image 2.5 Sunburst 質量平均分爲 7.38(滿分 10 分),但其廣告中僅 54.2% 可以直接投放。在八個模型中,可直接投放率介於 22.0% 至 54.2% 之間。所有模型都拿到相同的 59 份 brief,覆蓋四個品牌與五種語言,並由來自 OpenAI、Anthropic 和 Google 的 AI 評審組成盲測小組,逐條檢查每則廣告的文案是否精確、logo 與產品是否忠實、安全區以及是否存在捏造宣稱。這些均爲首次嘗試的數據,未啓用我們自己的複覈環節。 我們構建 TASTE BENCH,是因爲一旦你能度量品味(taste),你就能對它做爬山優化。它涵蓋圖片與視頻模型,你可以在 http://www.omneky.com/tastebench 查看每一條圖片廣告、brief、評審解釋與得分。
溯源
- 原作者
- omneky
- 發佈日期
- 2026-09-25
- 原始出處
- 查看原帖 ↗
- 收錄來源
- xianyu110/awesome-gpt-image2.5 · 2103352139838071038 · MIT