Opus 5.5 配班卓琴音乐讲图像 token 密度
作者用 Opus 5.5 生成一段视频，讲解 1080p 图等于 2961 个 token 及视觉预处理问题。
a quick peak into model & tokenization internals a 1080p picture is worth 2961 "words" (to glm-flash-5.3) every time you send in an image to a model, the harness makes a preprocessing decision how to resize and convert it into tokens good to internalize how token dense images and video are, especially at high frame rates/resolutions also think about compaction! what does it even mean to compact an image? you cant rewrite an image, the harness has to pick which to keep in context, which to pass references to, which to ignore all of this is a harness design decision when you see really low eval scores on visual tasks, sometimes models are blind because preprocessing distorts fine-grained image details. sometimes they just suck at perception. in both cases it's worth looking at what they actually "see" this time opus 5.5 set the vid to banjo music