我发出的编辑提示非常简单：

The edit prompt I asked was very simple: "Replace the kid on the cliff with Kai (character sheet reference image) and keep same acting posture as he is looking towards the valley. Fix some of the small people scale who look bigger than their houses down in the valley." Here are the results. Image 1: Original MidJourney render. Image 2: NanoBanana Flash. Image 3: NanoBanana Pro. Same edit. Same prompt. Different model. 👉Flash is great for speed. But if you're preparing keyframes for Veo, Seedance, Kling, Higgsfield, or any cinematic workflow where every frame matters... ...don't use Flash. 👉The Pro model preserves noticeably more of the original composition, lighting, textures, and material detail. That said... Even Pro still degrades a surprising amount of what MidJourney created. 😅 - Fine textures disappear. - Materials become softer. - Small design decisions get simplified. - The image slowly drifts away from the original art direction. 👉GPT Image 2...well see for yourself...I'm not impressed. Right now, I honestly don't think there's a model that can make significant edits while perfectly preserving everything that made the original image great. It's currently the weakest link in the AI filmmaking pipeline. Hopefully not for long.

我发出的编辑提示非常简单： "把悬崖上的小孩替换成 Kai（角色设定参考图），并保持他望向山谷时的同一表演姿态。修正山谷中一些看起来比自家房子还大的小人比例问题。" 以下是结果。 图像 1： 原始 MidJourney 渲染图。 图像 2： NanoBanana Flash。 图像 3： NanoBanana Pro。 同样的编辑。同样的提示。不同的模型。 👉Flash 胜在速度。 但如果你正在为 Veo、Seedance、Kling、Higgsfield 或任何电影级工作流准备关键帧，每一帧都至关重要…… ……那就不要用 Flash。 👉Pro 模型在保留原始构图、光照、纹理和材质细节方面明显更出色。 话虽如此…… 即便是 Pro，依然会出乎意料地损失大量 MidJourney 原本的细节。😅 - 精细纹理消失。 - 材质变得柔化。 - 一些微小的设计决策被简化。 - 图像慢慢偏离原本的艺术方向。 👉GPT Image 2……你自己看吧……我不满意。 老实说，目前我不认为有哪个模型能在进行重大编辑的同时，完美保留让原图出色的所有元素。 它目前是 AI 电影制作流程中最薄弱的一环。 希望不会持续太久。