A directing stage for generated 3D worlds.为生成式 3D 世界打造的导演舞台。
Holorra places generated assets inside walkable 3D Gaussian Splatting worlds, takes them apart into semantic parts, rigs characters to move through the space, and records camera paths for AI video. Every stage runs on the Manycore Aholo SDKs.Holorra 把生成的资产放进可行走的 3D 高斯泼溅世界,按语义拆分为部件,为角色绑骨,让它们在空间中走动,并录制用于 AI 视频的镜头路径。每个环节都运行在 Manycore Aholo SDK 之上。

Text-to-3D returns a mesh. It floats in a grey void, it is fused into one shell, it cannot move, and there is no way to film it. Holorra treats the generated asset as the start of a scene.文生 3D 只返回一个网格:它悬浮在灰色虚空里,熔成一整块外壳,不能动,也没法拍。Holorra 把生成的资产当作一个场景的起点。
| Gap缺口 | Typical text-to-3D output常见的文生 3D 输出 | In Holorra在 Holorra 中 |
|---|---|---|
| Place空间 | A lone .glb on a grey grid, with no room, street or light.灰色网格上孤零零的一个 .glb,没有房间、街道,也没有灯光。 | Inside a walkable 3DGS world from SpatialGen or Reality, with voxel collision on floors and walls.置于 SpatialGen 或 Reality 生成的可行走 3DGS 世界中,地面和墙面带体素碰撞。 |
| Parts部件 | One fused shell. Doors do not open and nothing separates for printing.一整块熔合的外壳。门打不开,也无法拆件打印。 | Semantic sub-meshes from lux3d.partSplit that can be exploded, isolated and recolored.lux3d.partSplit 输出语义子网格,可做爆炸视图、单独显示和改色。 |
| Motion运动 | Static props; characters locked in a T-pose.道具静止不动,角色定格在 T-pose。 | Joints from a sentence via articulationAnimation; humanoidAutoRig plus 85 retargetable clips.一句话经 articulationAnimation 生成关节;humanoidAutoRig 自动绑骨,另有 85 个可重定向的动作片段。 |
| Camera镜头 | A turntable at best.最多是一个转台展示。 | Keyframed camera paths with shot prompts, exported as a LuxReal storyboard.带镜头提示词的关键帧镜头路径,导出为 LuxReal 分镜。 |
A generated asset should be the start of a scene, not the end of a pipeline.生成的资产应当是场景的起点,而不是流水线的终点。
Holorra chains SpatialGen or Reality, Lux3D and the Aholo Viewer so the same asset can be placed, opened up, animated and filmed without leaving the browser.Holorra 串联 SpatialGen 或 Reality、Lux3D 与 Aholo Viewer,同一个资产无需离开浏览器,就能完成摆放、拆解、动画和拍摄。
Each step maps to official SDK calls and hands its output to the next step on the same WebGL2 stage.每一步都对应官方 SDK 调用,并在同一个 WebGL2 舞台上把产出交给下一步。
Next.js App Router. @manycore/aholo-viewer is loaded client-only and runs splat decoding and sorting in Web Workers.Next.js App Router。@manycore/aholo-viewer 仅在客户端加载,高斯点的解码与排序在 Web Worker 中运行。
/api/aholo/world, lux3d, tasks, director, status. Each wraps the official SDK client; the key comes from AHOLO_API_KEY or the x-aholo-api-key header./api/aholo/world、lux3d、tasks、director、status。每个路由都封装官方 SDK 客户端;密钥来自 AHOLO_API_KEY 或 x-aholo-api-key 请求头。
@manycore/aholo-sdk-mcp in .mcp.json for Claude Code and Cursor, plus an in-app co-director that turns a brief into an ordered tool plan.在 .mcp.json 中为 Claude Code 和 Cursor 配置 @manycore/aholo-sdk-mcp;应用内另有 AI 副导演,把一段需求转成有序的工具调用计划。
Live jobs take minutes, so the studio opens on four Gaussian Splatting scenes hosted on Manycore's CDN. The fifth came from one SpatialGen prompt through Holorra's own route.实时任务需要几分钟,所以创作台一打开就是托管在 Manycore CDN 上的四个高斯泼溅场景。第五个由一条 SpatialGen 提示词经 Holorra 自己的路由生成。


| World世界 | Format格式 | Splats高斯点 | Navigation漫游方式 |
|---|---|---|---|
| Designer Smart Loft设计师智能公寓 | Direct .spz直接加载 .spz | single file单文件 | Voxel walk · 336×112×296 · 5 cm体素碰撞行走 · 336×112×296 · 5 cm |
| Juguo Urban Plaza聚果城市广场 | Chunk-LOD + .ply sky分块 LOD + .ply 天空 | 41.8M · 5 levels · 528 chunks41.8M · 5 级 · 528 个分块 | Voxel walk · 1392×332×2336 · 20 cm体素碰撞行走 · 1392×332×2336 · 20 cm |
| Grand Transit Concourse交通枢纽大厅 | Chunk-LOD + .sog env分块 LOD + .sog 环境 | 32.7M · 5 levels · SH degree 232.7M · 5 级 · SH 2 阶 | Free camera自由视角 |
| Object Turntable物体转台 | Direct .sog直接加载 .sog | single object capture单个物体采集 | Free camera自由视角 |
| Coffee Bar · SpatialGen (ours)咖啡吧 · SpatialGen(我们生成) | Direct .spz, Z-up直接加载 .spz,Z 轴向上 | 266,895 · 625 s266,895 · 625 s | Fixed floor height固定地面高度 |
Splat totals and grid sizes are read from each world's lod-meta.json and voxel metadata. Chunk-LOD worlds stream by camera distance under a fixed on-screen splat budget.高斯点总数和网格尺寸读取自各世界的 lod-meta.json 与体素元数据。分块 LOD 世界按相机距离流式加载,屏幕上的高斯点数量控制在固定预算内。
Both meshes came from one text prompt through Holorra's own route. G1-Turbo returned a usable draft in 151 s; G1 took 343 s and embedded a full PBR texture set. The studio drafts with Turbo and spends G1 only on the hero asset.两个网格都由同一条文本提示词经 Holorra 自己的路由生成。G1-Turbo 用 151 s 返回可用的草稿;G1 用时 343 s,并内嵌整套 PBR 贴图。创作台先用 Turbo 打草稿,只为主角资产动用 G1。


| G1-Turbo | G1 | |
|---|---|---|
| Wall clock实测耗时 | 151 s | 343 s |
| Triangles三角面数 | 252,226 | 229,788 |
| Texture maps贴图数 | 2 | 3 (PBR) |
| GLB sizeGLB 大小 | 17.4 MB | 14.6 MB |
| Role用途 | draft, placement草稿、摆放 | hero, partSplit input主角资产、partSplit 输入 |
Prompt (both engines)提示词(两个引擎共用)“A modular espresso machine with a brushed titanium body, a hinged brewing hood on top, a sliding capsule drawer at the front and a rotary dial on the side. Product design, clean hard-surface geometry.”
Times are submit-to-success as seen by a 15 s poller. Geometry read from the returned GLBs.耗时为从提交到成功的时间,由 15 s 间隔的轮询记录。几何数据读取自返回的 GLB。
partSplit segmented the G1 mesh into 4 parts with the same 229,788 triangles. articulationAnimation turned one sentence into a URDF with exactly the two joints it asked for.partSplit 将 G1 网格拆成 4 个部件,三角面数不变,仍为 229,788。articulationAnimation 把一句话变成 URDF,恰好生成句中要求的两个关节。



| Part部件 | Our reading我们的解读 | Tris三角面 |
|---|---|---|
| part_0 | Main body主体 | 95,310 |
| part_1 | Side dial侧面旋钮 | 1,489 |
| part_2 | Top section顶部 | 59,340 |
| part_3 | Outer shell外壳 | 73,649 |
| URDF jointURDF 关节 | Type类型 | Range范围 |
|---|---|---|
| hood_top | revolute | 0° – 90° |
| drawer_front | prismatic | 0 – 0.233 |
| drawer_handle | fixed | — |
The first character was refused by the auto-rig. Regenerated in a strict T-pose it rigged in 81 s, and one retarget job mapped eight official clips onto its 52-joint skeleton.第一个角色被自动绑骨拒绝。改用严格 T-pose 重新生成后,81 s 即完成绑骨;一次动画重定向任务就把 8 个官方动作片段映射到它含 52 个关节的骨架上。



| Retargeted clip重定向动作片段 | In GLBGLB 内名称 | Length时长 |
|---|---|---|
| Idle_Loop | animation_1 | 2.5 s |
| Walk_Loop | animation_2 | 1.33 s |
| Jog_Fwd_Loop | animation_3 | 0.93 s |
| Interact | animation_4 | 2 s |
| Chest_Open | animation_5 | 1.37 s |
| Yes | animation_6 | 2.5 s |
| Dance_Loop | animation_7 | 1 s |
| Sword_Attack | animation_8 | 1.53 s |
The retarget GLB names its clips animation_1…8; mapped here in the order requested. All 85 official clips are selectable in the studio.重定向 GLB 把动作片段命名为 animation_1…8,此处按请求顺序对应。全部 85 个官方动作片段都可在创作台中选用。
Scored 70% by the expert panel and 30% by the community panel.评分构成:专家评审占 70%,社区评审占 30%。
| Criterion评分项 | Weight权重 | In Holorra在 Holorra 中 |
|---|---|---|
| Use of Manycore toolsManycore 工具使用 | 30% | The stage runs on @manycore/aholo-viewer (direct splats, chunk-LOD, GLTF skinning). Server routes under /api/aholo call the official World, Lux3D and Asset clients: all 10 Lux3D methods plus SpatialGen and Reality. Eleven live jobs ran through those routes: ten Lux3D (textTo3d ×3, partSplit, articulationAnimation, humanoidAutoRig ×2, humanoidAnimationRetarget ×2) and one SpatialGen world, now world 05 in the studio; every Lux3D result is inspectable on /lab. @manycore/aholo-sdk-mcp is pre-wired for agents.舞台基于 @manycore/aholo-viewer 运行(直接加载高斯点、分块 LOD、GLTF 蒙皮)。/api/aholo 下的服务端路由调用官方 World、Lux3D 和 Asset 客户端,覆盖全部 10 个 Lux3D 方法以及 SpatialGen 和 Reality。经这些路由实际运行了 11 个任务:10 个 Lux3D 任务(textTo3d ×3、partSplit、articulationAnimation、humanoidAutoRig ×2、humanoidAnimationRetarget ×2)和 1 个 SpatialGen 世界,即创作台中的世界 05;每个 Lux3D 结果都可在 /lab 查看。@manycore/aholo-sdk-mcp 已为智能体预先配置好。 |
| Fun and engagement趣味性与参与感 | 25% | Press Play for a one-minute scene quest: walk the Lux3D actor to the articulated espresso machine, brew (the URDF joints open in a camera close-up), frame the shot, celebrate. Text to 3D results drop into the world when ready, phones get a joystick, and Expand gives the render the full screen. Opens with no generation wait.进入游玩模式即开始一分钟的场景任务:操控 Lux3D 角色走到关节化的意式咖啡机前冲咖啡(镜头特写下 URDF 关节逐一打开),构图取景,然后庆祝。文生 3D 结果完成后会投放到场景中,手机端提供虚拟摇杆,展开后渲染画面占满全屏。打开即玩,无需等待生成。 |
| Innovation创新性 | 20% | Radiance-field worlds and part-split, articulated, rigged meshes share one stage, with a camera director that hands shots to LuxReal and an MCP co-director that plans tool calls from plain language.辐射场世界与经过部件拆分、关节化和绑骨的网格同处一个舞台;镜头导演把镜头交给 LuxReal,MCP AI 副导演则根据自然语言规划工具调用。 |
| Completeness完成度 | 20% | Studio, landing page, Lux3D bench, MCP hub, a 13-page /docs section and this report ship in one Next.js app, with bring-your-own-key support (x-aholo-api-key, COM / CN regions) and five export formats.创作台、首页、Lux3D 测试台、MCP 中心、13 页的 /docs 文档和本报告都集成在同一个 Next.js 应用中,支持自带 API 密钥(x-aholo-api-key,COM / CN 区域)和 5 种导出格式。 |
| Reach传播度 | 5% | Copyable .mcp.json for Claude Code and Cursor, a share-ready landing page, and the #Manycorehackathon tag on every page.可直接复制的 .mcp.json(适用于 Claude Code 和 Cursor)、便于分享的首页,且每个页面都带有 #Manycorehackathon 标签。 |