Skip to content

Model-To-VRM v0.1.4

Choose a tag to compare

@github-actions github-actions released this 09 Oct 10:14
· 38 commits to main since this release

新增 / New

  • 同时转换多个:几个文件、unitypackage 里的几套服装、压缩包里的几个模型同时转,快一倍以上(Marycia 整包 12.0 秒 → 3.8 秒)。「输出」页可选同时转几个:自动 / 1 / 2 / 4;「自动」按 CPU 核数和空闲内存定,每个任务要多少内存按文件大小估。某个模型失败不影响其它的;「停止」时已经在转的转完,没开始的不再开始
    Several files, a package's outfits or an archive's models convert in parallel (Marycia package 12.0 → 3.8 s); Auto / 1 / 2 / 4 on the Output tab, Auto sizes each job's memory from its file size; one failing model doesn't stop the others
  • 只转换人物(「VRM」页,默认开):模型包里常带着的道具(剑、椅子这类没有人形骨架的)和只有衣服 / 配件的模型自动跳过,日志里写明原因,文件列表显示「已跳过」。「只转 PMX」时不挑;关掉开关就照常全部转换
    "Characters only" (VRM tab, on by default): props without a humanoid skeleton and outfit-only models are skipped with the reason in the log; "PMX only" keeps everything

改进 / Improved

  • 转换更快:贴图只解码一次、多线程处理(Pillow),顶点权重只算一次,unitypackage 一遍解完。爱弥斯(13 万顶点)6.5 秒 → 1.9 秒,Marycia 的一个 FBX 5.1 秒 → 1.9 秒,TGA 贴图的 unitychan 1.8 秒 → 0.65 秒;输出和以前完全一样
    Faster: textures decoded once and in threads, weights computed once, one-pass unpack (Aemeath 6.5 → 1.9 s, Marycia FBX 5.1 → 1.9 s); outputs unchanged
  • 预览和游戏里更省内存:模型按表情分成几个网格(像 VRoid 的 Face / Body / Hair),每个网格只带动得到它的表情。三维软件和游戏会把表情展开成「顶点数 × 表情数」的完整数组,以前身体、衣服也要陪着脸带上全部表情:爱弥斯这一项 268 MB → 12 MB。表情效果不变,文件也小了一点
    Meshes split by expression like VRoid, so viewers and games don't expand every morph over the whole body (Aemeath: 268 MB → 12 MB of morph data); expressions look the same
  • 转换时的内存峰值也低了一些(爱弥斯 600 MB → 510 MB)
    Lower peak memory while converting
  • BMP / TGA / DDS 贴图用 Pillow 转换,更快,8 位 BMP 也能读了
    BMP / TGA / DDS textures converted with Pillow: faster, 8-bit BMPs now load
  • 压缩包里不同文件夹的同名模型输出成 _2、_3,不再互相覆盖
    Same-named models in different folders of an archive get _2 / _3 instead of overwriting each other