Latent Space is the GPU-render visualization of a machine-learning embedding manifold: a luminous violet-to-blue point cloud suspended in a black render void. It borrows the visual grammar of t-SNE and UMAP scatter plots, where high-dimensional data collapses into glowing clusters threaded with Perlin and Gaussian noise.
This system renders the internals of generative models as scientific imagery — embeddings, dimensions, and manifolds drawn in additive glow on black. It says, plainly: this is what a model sees inside itself.
「潜空间」是机器学习嵌入流形的 GPU 渲染可视化——在一片漆黑的渲染虚空中，悬浮着一团由紫到蓝、熠熠发光的点云。它借用 t-SNE 与 UMAP 散点图的视觉语法：高维数据在降维后坍缩成发光的聚类，其间穿插着 Perlin 与高斯噪声。
这套设计语言把生成式 AI 的内部结构呈现为科学影像——嵌入、维度与流形以加色辉光绘于黑底之上。它直白地宣告：这就是模型在自身内部所「看见」的样子。绝不使用米色或纯白，底色必须是黑色的渲染虚空。