Comparing Four Models with an Abstract Visual Prompt
Examine how different models interpret an open-ended request for an unprecedented, striking image. The case explains why concrete subjects, scale, lighting and composition make comparisons more informative.
English translation of the public case notes.
Comparison case examining how four models respond to the same highly abstract English prompt. A grand, exaggerated visual request gives models considerable freedom, revealing differences in instruction following, default aesthetics and their interpretation of a striking image. Input: one unchanged abstract prompt with side-by-side model outputs. Goal: observe interpretations of an unprecedented, extremely impressive image. The prompt sets an ambitious overall effect but leaves subject, setting, composition, materials and lighting unspecified, so results depend heavily on each model's own defaults. Avoid relying entirely on stacked superlatives. They can produce generic spectacles that are difficult to reproduce and provide weak evidence for evaluating a particular capability. Practical controls: • Translate the desired impact into concrete elements: scale, subject, surroundings, light source and viewpoint. • Keep the original prompt unchanged during a comparison and examine default composition and style. • For greater control, add a specific medium and structure, such as surreal architecture, cosmic scale, central composition or macro material detail.
Original case notes (Chinese)
🧯 翻车点观察｜四模型同题想象力对照 实验类型：文生图 / 图像编辑。 这次主要验证的是：这个案例适合看抽象夸张提示词在不同模型中的分化：主题越宏大，模型越容易各自发挥。可用于测试指令遵循、审美默认值与“惊艳感”生成边界。 实验条件： 类型：对照测试 输入：同一条高度抽象的英文提示词，以及多个模型的并排输出 目标：观察不同模型如何理解“前所未见、极度震撼”的开放式视觉要求 关键词：视觉震撼 / 前所未见 / 抽象宏大 / 模型对照 / 审美默认值 / 指令发散 重点：提示词先给“最震撼、无人见过”的总体目标，但缺少主体、场景、构图、材质与光线，结果主要依赖模型自身审美补全 约束：应避免只用夸张形容词堆叠；否则容易生成泛化的奇观画面，难以复现，也不利于判断具体能力 💡 结果最说明问题的三点 • 把“震撼”拆成可执行元素，例如尺度、主体、环境、光源和观看视角 • 对照测试时保持原提示词不变，重点观察模型默认构图和风格倾向 • 若要提升可控性，可追加明确媒介与画面结构，如超现实建筑、宇宙尺度、中心构图或微距材质 #GPTImage2 #OpenAI #AI绘画 #模型对照 #超现实奇观 #抽象提示词