Opus 5.5 加 GPT Image 2.5 做 MV 的失败复盘
作者公开用 Opus 5.5 高推理模式、本地跑 GPT Image 2.5 和 MiniMax H3 生成 MV 的过程，耗时 8 小时仍有连贯性和快速运动画质问题。
Again on the theme of building in public. This is to show that all of these Frontier models are not infallible. This is the work of Opus 5.5 on high reasoning. This took over 8 hours to complete. Most of the time-consuming work was generating images and video using GPT Image 2.5 and MiniMax H3 on my local hardware. There are continuity problems, issues with degradation due to rapid movement, etc. The bigger thing is that Opus decided to take the artist that created this song and make it rather literal for the theme. I am going to redo it and hopefully this will improve how I address these types of videos going forward, as this is going to be another option inside of Sound / Vision.