RFD 1102 details: the task table, splat status, architecture, and further reading

Task types, New Task panel

Task	API feature	Example models (DGX, June 2026)
Text to 3D	`text_to_textured_mesh`	TRELLIS
Image to 3D	`image_to_textured_mesh`	TRELLIS.2 (recommended), Pixal3D (PBR), Hunyuan3D-2.1
Image to Raw Mesh	`image_to_raw_mesh`	Hunyuan3D-2.1, UltraShape
Mesh painting (text or image)	`text_mesh_painting` / `image_mesh_painting`	TRELLIS.2, Hunyuan
Mesh segmentation	`mesh_segmentation`	P3-SAM
Mesh retopology	`mesh_retopology`	AutoRemesher (default), Instant Meshes, Trimesh Decimate
Mesh UV unwrapping	`uv_unwrapping`	xatlas
Mesh editing (text or image)	`text_mesh_editing` / `image_mesh_editing`	VoxHammer
Auto rigging	`auto_rig`	SkinTokens (full GLB, recommended), UniRig (template VRM)
Text to Motion (Kimodo)	`text_to_motion`	Kimodo SOMA-RP-v1.1, into studio motion JSON, into VRM/rigged-GLB playback
Image to Gaussian Splat	`image_to_splat`	TripoSplat (1 photo), WorldMirror 2.0 (2+), COLMAP (3+)
Image to World	`image_to_world`	`weftspun_image_to_world` (a splat environment, plus optional TRELLIS.2 props)
Avatar from Image	a client pipeline	TRELLIS.2 mesh, into a UniRig template rig, into a GLB
Avatar From Photo	client only	AvatarSDK, not `3DAIGC-API`

Also shipped, client plus API: multi-image input, a primary photo plus up to seven references, on splat, world, and avatar tasks (RFD 1094); Publish RP1/OMB validate, sending a mesh job to the spatial fabric through the MSF Map Service (RFD 1100); Kimodo text-to-motion, an animation-bar prompt into a SOMA motion job into viewport playback on a VRM or a rigged GLB (KimodoMotionPromptBar.jsx, kimodoMotionLoader.js).

Not in the UI: “Part completion” (legacy upstream docs only). License-blocked on a commercial tier: PartField, PartPacker, FastMesh; see 3DAIGC-API’s own MODEL_LICENSES.md.

Gaussian splats (3DGS)

Splats live inside this app, the same SceneManager viewport as VRM and mesh workflows, not a separate product. Generation runs on the DGX, through 3DAIGC-API; viewing uses Spark.js (sparkSplatManager.js, @sparkjsdev/spark) in the main Three.js scene.

Shipped today

Capability	Client	API (DGX)
Splat preview in the viewport	A `SplatMesh`, alongside VRM and meshes	`POST /api/v1/splat-generation/image-to-splat`
One photo to a splat	Task Manager’s multi-select, one primary photo	TripoSplat
Two or more photos to a splat	The same UI; mark the best front view as primary	WorldMirror 2.0 (COLMAP as the fallback, at three or more)
World package load	World Library, plus `worldSceneLoader.js`	`POST /api/v1/world-generation/image-to-world` (`weftspun_image_to_world`)
A walked, XR environment scan	Task Manager’s “Environment Scan”	`POST /api/v1/world-generation/environment-scan` (LingBot-Map)
Env-scan Phase A, into Spark	Automatic, when `refine_to_3dgs` is set	Isotropic Gaussians, from a point cloud
Env-scan Phase B train	Separate, or `train_3dgs: true`	`POST /train-3dgs` / `env_scan_gsplat_train`, 7 or 10,000 steps
Avatar plus optional splat	“Avatar from Image,” plus a “Gaussian splat preview” checkbox	TRELLIS.2 mesh, plus a UniRig template rig, plus an optional TripoSplat
Multi-image uploads	`multiImageInput.js`, on splat, world, and avatar tasks	`image_file_id` plus `reference_image_file_ids`, up to eight images total

Task types in the New Task panel: “Image to Gaussian Splat,” “Image to World (splat + props),” “Environment Scan.”

Not done yet

Splat-only world RP1: World Library’s own RP1 needs mesh props in the manifest; a pure Gaussian environment needs a prop-generation path before it can publish to OMB at all (RFD 1100 gives why). Env-scan x402 SKUs: Phase A versus Phase B billing, plus a frame-budget upsell, is roadmap work, not shipped. Gaussian-VRM/RGBAvatar body pipelines: scan-based full-body avatars, and the highest-fidelity head-attachment path, are separate from the viewport’s own TripoSplat preview, not the same code path as image-to-splat.

Where it lives, architecture

[DGX 3DAIGC-API]  TripoSplat, WorldMirror/COLMAP, image-to-world, LingBot env-scan A/B, avatar mesh/rig jobs
       |
[this project /]  SceneManager: one renderer, one WebXR session, VRM + tools
       |                  SparkRenderer + SplatMesh (LingBot: orientationMode none)

/xr stays an IWSDK lab for grab and locomotion regression (RFD 1090). The main app (/) runs IWSDK Option A on SceneManager: distance and proximity grab (trigger), a grip that opens a context menu or pans, and thumbstick locomotion, alongside loaded splat worlds and a VRM, in the same session.

Application feature summary

Rendering modes: Solid, Rendered, Wireframe, Skeleton, Part Colorize. File formats: GLB, GLTF, OBJ, FBX, DAE, STL, VRM; images: JPG, PNG, BMP, TGA, for image-to-3D workflows.

WebXR: VR with a floor-anchored, virtual-sky scene; AR with pass-through transparency; Samsung Galaxy XR (Chrome WebXR) as the primary on-device target; bounded-floor, local-floor, local, and viewer reference spaces (RFD 1108 gives the implementation); WebXR expression tracking when the browser exposes it, a native face relay through a companion APK when it does not (RFD 1096, RFD 1082).

Also shipped: WebGPU rendering with an automatic WebGL fallback; SSAO, Bloom, and FXAA post-processing; positional (spatial) audio; a Core3D integration for a model and materials library; a shared, mode-detecting 3D viewer (Universal3DViewer).

Avatar and VRM: a trait-based VRM character system, with a soulbound base body and equippable clothing, hair, and accessory layers; drag-and-drop texture and model overrides; Mixamo integration and Kimodo text-to-motion for animation; blend shapes, lip sync, eye tracking, and automatic blinking; an optimized VRM export with texture atlasing and mesh merging; batch VRM generation from a manifest; one-click optimization down to a single draw call.

Key components

Avatar and viewport: Scene3D, TaskManager, FileUpload, RenderModeSelector, APIStatus, SceneManager (the WebGL-based core), Shared3DViewer, Universal3DViewer, Core3DViewer, Core3DPanel, Core3DContext, Core3DService.

Character and animation: CharacterManager, AnimationManager, BlinkManager, LookAtManager, EmotionManager. State: taskStore, sceneStore, SceneContext, Core3DContext.

This project began as Open3DStudio, a WebGL-only foundation with basic 3D AIGC workflows, task management, and file import/export. The Three.js WebGPU and WebXR migration guide documents the move from that foundation to the current rendering stack.