圆珠笔草图硬刚成片：Image2 vs Flare/Sunburst

提示词

A ballpoint scribble went in as the only image. Three finished photographs came out.

GPT Image 2 vs 2.5, specimen 03: sketch to image. The input was a crude composition diagram in ballpoint, shapes only, with handwritten notes such as 5 ARCHES and KEEP HIM SMALL. The prompt told each model to read it as a blueprint and follow the notes. Same prompt, one run each, first result kept. GPT Image 2 at High, 2.5 Sunburst and 2.5 Flare at Max, all 4K.

All three followed the diagram: five arches, a locomotive and four carriages heading left with steam trailing right, a small figure on a boulder at lower left with a bicycle, pine at the right edge, river from the lower right, sun upper right behind the hill. No handwriting made it into the picture, no paper texture either.

Where they part: GPT Image 2 turned the gorge into open moorland with a wide shallow river, and its mist under the arches is a few wisps near the water. Sunburst has the densest mist under the arches. Flare's mist sits mostly in the far valley.

Verdict, from one person going through the originals: GPT Image 2 and Flare look better. Sunburst has a lot of high-frequency noise, so the image looks dirty and less clear. That was at Max quality.

Method and what you pay in the reply.

中文译文

一张圆珠笔涂鸦作为唯一输入。输出三张成品照片。

GPT Image 2 vs 2.5，样本 03：草图到图像。输入是一幅用圆珠笔画的粗糙构图示意图，只有形状，并带有手写注释，例如 5 ARCHES 和 KEEP HIM SMALL。提示词要求每个模型把它当作蓝图来读，并遵循这些注释。相同提示词，各跑一次，保留第一个结果。GPT Image 2 使用 High，2.5 Sunburst 和 2.5 Flare 使用 Max，全部 4K。

三张都遵循了示意图：五道拱门；一辆机车和四节车厢朝左行驶，蒸汽向右拖曳；左下方一块巨石上有个小小的人影和一辆自行车；右边缘有松树；河流从右下方延伸而来；太阳在右上方山丘之后。画面中没有出现任何手写笔迹，也没有纸张纹理。

它们的差异在于：GPT Image 2 把峡谷变成了开阔的荒原，配一条宽阔的浅河，拱门下的薄雾只是水边几缕。Sunburst 拱门下的雾最浓密。Flare 的雾大多停留在远处的山谷中。

结论，来自一个人逐一查看原图：GPT Image 2 和 Flare 看起来更好。Sunburst 有大量高频噪点，因此画面看起来脏、不够清晰。而这还是在 Max 质量下。

方法与所需花费见回复。

溯源

原作者

FlyAIgh

发布日期

2026-09-16

原始出处

查看原帖 ↗

收录来源

xianyu110/awesome-gpt-image2.5 · 2100205158911361300 · MIT